Demystifying SharePoint Performance Management Part 9 – Don’t believe everything you R/W

This entry is part 9 of 11 in the series Perf
Send to Kindle

Hi and welcome to Part 9 (bloody hell… nine!) of my series on trying to demystify SharePoint performance management a bit. If by any chance you have been asked to provide some sizing information for your organisation and you are finding the resources online a bit overwhelming, this series is for you. If you have been a part of our varied journey so far, the last few posts have been all about Disk IO performance in the form of latency, IOPS and MBPS. In the last two articles, we have been learning about the different IO patterns that SQL Server is likely to utilise, as well as using the jackhammer known as the SQLIO utility, that is used to simulate those IO patterns on unsuspecting disk infrastructure.

Now just to set the scene for this post (and conveniently perform some product placement), I recently published a book called “The Heretics Guide to Best Practices”. Now being the author and all, I am going to suggest you buy it because it is a completely riveting read! :-).

Now apart from blatant product placement, the real reason I mention it is because one of the chapters is called “Myths, Memes and Methodologies”. In it, we examine why some ideas gain legitimacy, even though they are based on often completely dodgy foundations. I mention this here, because in terms of SQL disk IO sizing, something similar has happened with Microsoft’s published material on the topic. So the focus of this article is to finish off our discussion on understanding disk IO patterns, while lifting the lid on some of the inconsistencies in the material that that end up being repeated by SharePoint consultants as gospel to their unsuspecting disciples.

Now harking way back to part 1 to the notion of lead vs. lag indicators, our use of SQLIO thus far has essentially been used as a lead indicator. While SQLIO puts a real load on disk infrastructure and faithfully reports the resulting IOPS, latency and MBPS, the reality is it can never truly capture the nuances of a production SharePoint farm doing its thing. But in terms of a lead indicator that is okay. After all, a lead indicator by definition cannot guarantee an outcome. It can merely suggest that an outcome should be able to be met.

So while we are thinking about the lead indicator world view, some of you might have noticed that I have not yet made any suggestions what are the minimum conditions of satisfaction for disk infrastructure used to underpin SharePoint. This has been deliberate until now, because I felt that it was critical to understand the relationship between the size of a disk IO operation, and its effect on IOPS, latency and MBPS first. To that end, hopefully I have instilled a reflex in you where – if you are given an arbitrary latency, IOPS or MBPS figure that you have to meet – you immediately ask questions like, “What sort of IO patterns?” or “how large is the IO request typically going to be?” or “is the IO random or sequential?”

When whitepapers mislead…

Now we are about to get into one area where Microsoft’s published documentation is quite weak. Remember the 367 page “Capacity Planning for Microsoft SharePoint Server 2010” whitepaper? Starting at page 326, there is a section with the promising title of “Estimate Core Storage and IOPS needs” (this topic is also available separately as a technet article too). The problem is in despite that title, very little IOPS guidance actually is given. Instead the content in the section overwhelmingly speaks about estimating storage requirements. In fact the best you get is one explicit mention of IOPS in relation to the SharePoint Search service application which states the following:

The IOPS requirements for Search are significant.

  • For the Crawl database, search requires from 3,500 to 7,000 IOPS.
  • For the Property database, search requires 2,000 IOPS.

Note: For the purpose of the rest of this article, lets add the above figures together and simply say between 5,500 to 9000 IOPS for search.

Do you see the problem here? This is simply an arbitrary IOPS figure with no guidance as to the IO patterns underpinning it. What about latency or the IO request size that you need to assume? Unfortunately, no guidance is given for these questions which makes this quoted figure not overly helpful. Plus, as you will soon see below, Microsoft seemingly contradict themselves elsewhere in the same whitepaper…

So what are good numbers to use?

In the absence of any hard data, the best way to deal with storage requirements is to think in terms of lead indicators. Indicators from a lead point of view, can be framed as targets – something to aim for. Targets then can be broken down into different categories ranging from “cover your arse” to “above and beyond”:

  • Aspired target: The “this would be bloody fantastic if we could get there” target.
  • Agreed target: The “this is what we are going to deliver no matter what” target.
  • Minimum Condition of Satisfaction (MCOS) target: The “If we don’t achieve this we may as well pack up and go home” target.

So given these sorts of targets, what should the disk IO performance targets for SharePoint be? To work this out, we can utilise information already out there. Well…that is, we could if the information out there wasn’t so disparate and disconnected. So unfortunately, it takes some digging to you can find what you need.

Our first point of call in this regard is indeed Microsoft and the very same capacity planning and configuration guide that I criticised earlier for poorly dealing with IOPS. Hidden in the bowls of that document, the following statement is made on page 334 (emphasis mine):

Any storage architecture must support your availability needs and perform adequately in IOPS and latency. To be supported, the system must consistently return the first byte of data within 20 milliseconds.

So the way I look at it, a 20ms latency should be our MCOS target (see the explanation above for MCOS). If we consistently do worse than this, then we do not have a lot of assurance about the scalability of the disk IO subsystem being used for SharePoint. But like the arbitrary IOPS figure quoted in the previous section, I wonder if readers have spotted the problem with specifying this latency figure alone?

In both cases, don’t forget the almost symbiotic type of relationships between IO size, IOPS and latency. If we assumed that all IO operations were small (for example SQL’s page size of 8KB) then we could likely stay way under the 20ms limit with a more modest disk infrastructure. But to sustain the same latency with a larger IO size would require a faster disk subsystem. Why? Well as we discussed in part 6, if the size of the IO writes are larger, such as 64KB, then latency will go up because servicing larger requests takes longer than smaller ones. Therefore, if we were to assume a larger IO size, we would need more/faster disks to be able to meet the same 20ms latency KPI.

So what disk IO size should we assume to give context to a latency figure? Some insight can be found back in part 6, when we examined SQL IO characteristics and established that it will likely be much more varied than SQL’s base IO unit of 8KB pages. My suggestion therefore, is to test 8KB but also ensure that 64KB can meet the latency target. This is because 64KB represents a reasonable average size between the 8KB to 256KB range most SQL Server’s IO operations will fall within. Thus, if a SQLIO test using random read/writes at 64K indicates more than 20ms latency consistently, then you should probably ask your storage people to take another look at it.

By the way, if you really want to give your storage guys a challenge, keep jacking up the IO size!

What about aspired latency targets?

So if you are cool with the notion that the minimum condition of satisfaction for a random IO test using 64K size should be less than 20ms latency, what about aiming higher with agreed or aspirational targets?

Luckily for all of us, we can once again stand on the shoulders of giants. In this case, the Bob Duffy indirectly answers this question by providing what he considers to be the indicators for optimal SQL Server performance in general. In an excellent article with the rather appropriate title of “How to Specify SQL Storage Requirements to your SAN Dude” Bob makes the following recommendations:

  • SQL Data files must have a response time averaging about 8ms and a maximum response time of around 20ms using 64k size IOs and that are random in nature
  • SQL Log Files must have a write response time averaging from 1-5ms. use 64k IO size and are sequential in nature

The nice thing about specifying a target or benchmark like this, is that you are able to sidestep discussions on RAID levels, stripe sizes and many other things that SAN nerds find interesting. We keep things focused on the lead indicators and in effect state “If you can meet these figures, configure it any way you like.” This gives the SAN guys the freedom to do their job, while giving you an indicator that can give you confidence in the disk infrastructure. So if we were to distil the figures above into lead indicator targets for storage gurus, it might look something like this:

  • MCOS target: Less than 20ms latency for random IO requests of 64KB
  • Agreed target: Average 8ms latency for random IO requests of 64KB with no more than 20ms max latency. Less than 5ms latency for sequential log IO
  • Aspired target: No more than 8ms latency for random SQL IO requests of 64KB and average of 1ms latency sequential log IO with max never going above 5ms

Now in the ProData article, Bob made a slightly tongue in cheek point that sums up the above thinking really well, as well as giving insight to a critical aspect we have not considered so far…

Nowadays most SQL consultants try and not talk about RAID types and types of disk, it can be best to leave that up to the storage guys. If the storage team can meet my requirement for 5,000 random 64k read/write IOPs at 8ms latency by using 50 old SATA drives at 5,400 rpm in RAID 5 then knock yourself out – I’m happy. Well maybe I’m happy till we have that chat about Service Level requirements during a disk degrade event but that’s a different story…

If you look closely at Bob’s quote, you will see that he has also specified the last critical variable in the mix. Bob’s mention of “5000 random 64k read/write IOPS” is in reference to another point he makes. Without an IOPS figure to work from, the targets we have come up with are effectively meaningless. Quoting Bob:

The main thing to specify apart form your latency requirement is the throughput (IOPs). It is no good meeting the 8ms target for 100 IOPs and then finding your workloads needs 5,000 IOPs. You wont be able to meet the 8ms target!!

Consider it this way… a SharePoint site that services 100,000 users, will process a lot more IO requests than a site that services 10 users. With the latter, it is quite likely that the latency targets we have been talking about (even the aspirational ones) would be pretty easy to meet with a single disk. (To hark back to our shopping centre metaphor, one check out operator is all that is needed at a corner store, whereas many are needed at the supermarket). This is presumably why Bob has used a figure like 5000 IOPS for his post. It is probably a figure that conveniently represents some fairly heavy disk usage. But it does beg two question:

  • How much IOPS should we use to simulate SharePoint IOPS?
  • In the absence of anything else, perhaps 5000 IOPS is a good figure to go with?

Don’t believe all you read…

Now if you go back and read the start of this post, you will recall I mentioned that Microsoft stated some IOPS figures for the SharePoint search application databases ranged between 5,500 to 9000. That would indicate that Bob’s base figure of 5000 is a bit low, especially given that SharePoint has many other components beyond search that have not been taken into account. So to put Bob’s 5000 IOPS figure in perspective, let’s re-examine Microsoft’s trusty capacity planning whitepaper. One of the great things about this document is that Microsoft detailed the performance stats of a typical day in the life of their internal SharePoint environments. Since Microsoft are so large, they have different SharePoint farms for different collaborative scenarios. The scenarios they covered were:

  1. Enterprise Intranet environment (also described as published intranet). In this scenario, employees view content like news, technical articles, employee profiles, documentation, and training resources. It is also the place where all search queries are performed for all of the other the SharePoint environments within the company.
  2. Enterprise intranet collaboration environment (also described as intranet collaboration). In this scenario, is where important team sites and publishing portals are housed. They are typically used for enterprise collaboration, organizations, teams, and projects. Sites created in this environment are used as communication portals, applications for business solutions, and general collaboration. No custom code runs in these sites.
  3. Departmental Collaboration environment. In this scenario, employees use this environment to track projects, collaborate on documents, and share information within their department.
  4. Social Collaboration Environment. This is the My Sites scenario. These connect employees with one another and the information that they need. Employees use this environment to present personal information such as areas of expertise, past projects, and colleagues to the wider organization. The environment also hosts personal sites and documents for viewing, editing, and collaboration.

Now reading about these scenarios is highly interesting and Microsoft provides some nice nuggets of information that we will use in a future post. But for now I will stick purely to a disk IOPS perspective. To that end, below are a few fun-filled facts about the number of users in each of the four scenarios:

  1. Enterprise Intranet environment:  33580 unique users per day, with an average of 172 concurrent and a peak concurrency of 376 users.
  2. Enterprise intranet collaboration environment: 69702 unique users per day, with an average of 420 concurrent users and a peak concurrency of 1433 users
  3. Departmental Collaboration environment. 9186 unique users per day, with an average of 189 concurrent users and a peak concurrency of 322 users
  4. Social Collaboration Environment. 69814 unique users per day, with an average of 639 concurrent users and a peak concurrency of 1186 users

So now you have a sense of the size of these scenarios and as an added bonus, gotten a glimpse into the difference that usage patterns can make. For example: social collaboration and enterprise collaboration have similar number of unique users but social has more average concurrency but less peak. But what about IOPS?

In the document, IOPS is split into reads per second and writes per second, so I added them to estimate IOPS. The results are rather surprising…

Metric

Social Collaboration

Departmental Collaboration

Published intranet

Intranet Collaboration

Unique visitors

69814

9186

33580

69702

Average concurrent

639

189

172

420

Max concurrent

1186

322

376

1433

IOPS

941

74

409.66

409.66

Now while it might be tempting to ponder why social collaboration has over double the IOPS, yet half the concurrency of enterprise intranet collaboration, we are not going to worry about here. Besides, we actually covered some of it already when we used logparser to get insights of usage patterns. What I will instead do is draw your attention to is the fact that that none of the IOPS scenarios come anywhere near the 5000 IOPS figures cited by ProData or Microsoft’s 5500-9000 IOPS cited for search (in the very same capacity planning document I might add!)

So something is amiss. If an organisation the size of Microsoft can have almost 70000 unique users per day, with a peak concurrency of 1433 users and only total 410 IOPS, then where the hell did the 5500-9000 IOPS figure for search alone come from? Even if you take the scenario with the highest IOPS (the Social collaboration scenario with 941 IOPS), that’s still less than one fifth 5500 IOPS which was at the low end of the search IOPS figure.

Now I am also suspicious that two different case studies have the exact same IOPS figure. If you compare the “published intranet” scenario with the “intranet collaboration” scenario, one has half the visitors, yet both have precisely the same IOPS (right down to decimal places). That seems highly unlikely to me and I suggest that a mistake has been made. Given the intranet collaboration has the highest max concurrency figure, I would have expected IOPS to be a higher than it is. Hmmm…

What can we take away from this? For one, the capacity planning document could seriously do with a rewrite in this area. Secondly, I don’t have a lot of faith in those IOPS figures quoted (although I have more confidence in the case studies that the arbitrary figures specified for search).

So if we put aside the doubt created by the issues with the capacity planning guide, there is one really interesting fact that remains… none of the reported IOPS figures came anywhere near 5000 IOPS.

Insights from HP…

It turns out that Hewlett Packard also did some load testing of SharePoint 2010 (among other things) and published a whitepaper called the “HP performance and configuration guide for Microsoft SharePoint 2010“.  In this guide, they detail the results of a scenario they tested based on what they termed an “Enterprise Workload”. The guide covers definition of enterprise workload in loving detail, but the gist of it is that it covers the following areas:

  • Document Center (30% of operations) Check-out, download, upload and check-in documents
  • Team Sites – (20% of operations) work with calendars, discussions and documents
  • Portal SItes – (20% of operations) work with event, announcements and surveys
  • My Sites – (10% of operations) work with documents in personal documents library
  • Search – (20% of operations) Submit searches with random word or phrases

HP then simulates 500 concurrent users performing the actions above. In Table 13 of the report (page 28 of their document and reproduced below) , HP outline the performance and even break down the IO characteristics of each SharePoint database (which is really handy indeed). Adding up the last column of transfers/sec (which is essentially IOPS) we get a result of 1347.33 IOPS.

Thus we are still considerably under the 5000 IOPS that Bob Duffy suggests.

Conclusion…

Right! Remember our discussion above on MCOS, agreed and aspired targets? For an aspirational target, I think that we can reasonably use 5000 IOPS as a starting point for an enterprise organisation of Microsoft’s size. If we stick with 5000 IOPS, then my suggestion for an aspirational latency target would be:

  1. no more than 8ms latency for random SQL IO requests of 64KB
  2. average 1ms (and no more than 5ms max) latency of sequential log IO of 64KB

I think these figures are a pretty good test of a disk subsystem and think that Bob at ProData is therefore pretty close to the mark. Of course, you can use these figures to make your own judgement and adjust accordingly. Provided that you think of them as lead indicators that provide you a level of confidence in your disk infrastructure, you now have the tools and knowhow to run the tests too.

So if there was a moral of the story to this post, it would be to not believe everything you read and always verify espoused reality with actual reality via testing. On that note, the next post will finish off our examination of disk performance by going over 2 additional tools that I think are particularly good for testing assumptions. After that, we will be revisiting Microsoft’s case studies, as well as some findings, insights and recommendations from some additional lab scenarios that Microsoft conducted.

Thanks for reading

Paul Culmsee

www.sevensigma.com.au

 Digg  Facebook  StumbleUpon  Technorati  Deli.cio.us  Slashdot  Twitter  Sphinn  Mixx  Google  DZone 

No Tags

Send to Kindle

Demystifying SharePoint Performance Management Part 8 – More on SQL and SQLIO

This entry is part 8 of 11 in the series Perf
Send to Kindle

Well here we are at part 8 of my series on making SharePoint performance management that little bit easier to understand. What is interesting about this series is its timing. If by some minute chance that the marketing tsunami has passed you by at the time I write this, SharePoint 2013 public beta was released. Much is being made about its stated requirement of 24GB of RAM for a “Single server with a built-in database or single server that uses SQL Server”. While the reality is that requirements depends on what components that you are working with, this series of articles should be just as useful in relation to SharePoint 2013 as for any other version.

Now, if you have been following events thus far, we have been spending some time examining disk performance, as that is a very common area where a sub optimal configuration can result in a poor experience for users. In part 6, we looked at the relationship between the performance metrics of disk latency, IOPS and MBPS. We also touched on the IO characteristics (nerd speak for the manner in which something reads and writes to disk) of SQL Server and some SharePoint components. In the last post, we examined the windows performance counters that one would use to quickly monitor latency and IOPS in particular. We then finished off by taking a toe dip into the coolness of the SQLIO utility, that is a great tool for stress testing your storage infrastructure by pulverising it with different IO read and write patterns.

In this post, we will spend a bit of time taking SQLIO to the next step and I will show you how you can run a comprehensive disk infrastructure stress test. Luckily for the both of us, others have done the hard work for us and we can reap the benefits of their expertise and insights. First up however, I would like to kick things off by spending a little time showing you the relationship between SQLIO results and performance monitor counters. This helps to reinforce what the reported numbers mean.

Performance Monitor and SQLIO

In the previous post when we used Windows Performance Monitor, we plotted IOPS and Latency by watching the counters as they occurred in real-time. While this is nice for a quick analysis, nothing is actually stored for later analysis. Fortunately, performance monitor has the capability to run a trace and collect a much larger data set for a more detailed analysis later. So first up, lets use performance monitor to collect disk performance data while we run a SQLIO stress test. After the test has been run, we will then review the trace data and validate it against the results that SQLIO reports.

So go ahead and start up performance monitor (and consult part 7 of this series if you are unsure of how to do this). Looking at the top left of the performance manager, you should see several options listed under “Performance”. Click on “Data Collector Sets” and look for a sub menu called “User Defined”. Now right click on “User Defined” and choose “New –> Data Collector Set” as shown below:

image

This will start a wizard that will ask us to define what performance counters we are interested in and how often to sample performance. I have pasted screenshots of the sequence below (click to enlarge any particular one). First up we need to give a name to this collection of counters and as you can see below, I called mine “Disk IO Experiments”. Once we have given it a name, we have to choose the type of performance data we want to collect. Tick the “Performance counter” button and ensure the others are left unticked.

image  image

Next we need to pick what specific counters we need. We will use the same counters that we used in part 7, with the addition of two additional ones. To remind you of part 7, the counters we looked at were:

  • Avg. Disk sec/Read   – (measures latency by looking at how long in seconds, a read of data from the disk took)
  • Avg. Disk sec/Write  –  (measures latency by looking at how long in seconds, a write of data to the disk took)
  • Disk Reads/sec  –  (measures IOPS by looking at the rate of read operations on the disk per second)
  • Disk Writes/sec  – (measures IOPS by looking at the rate of write operations on the disk per second)

In addition to these counters, we will also add two more to the collector set

  • Avg. Disk Bytes/Read – (Measures size of each read request by reporting the number of bytes each used)
  • Avg. Disk Bytes/Write – (Measures size of each write request by reporting the number of bytes each)

We will use these counters to see if the size of the IO request than SQLIO uses is reported correctly.

Depending on your configuration, choose the PhysicalDisk or LogicalDisk  performance object (consult part 7 for the difference between PhysicalDisk and LogicalDisk). You will then find the performance counters I listed above. Before you do anything else, make sure that you pick the right disk or partition from the “Instances of selected object” section. We need to specifically pick the disk or partition that SQLIO is going to stress test. Now you select each of the aforementioned six performance counters and click the “Add” button. Finally, make sure that you pick the sample interval to be 1 second as shown below. This is really important because it makes it easy to compare to SQLIO which reports on a per second basis.

image  image

At this point you do not need to configure anything else, so click the “Finish” button, rather than the “next” button, and the collector should now be ready to go. It will not start by default, but since there is no fun in that, let’s collect some data. Right click on your shiny new data collector set and choose “Start”.

image  image

Once started, performance monitor is collecting the values of the six counters each and every second. Now let’s run a SQLIO command to give it something to measure. In this example, I am going to run SQLIO with random 8KB writes. But to make it interesting, I will use two threads and simulate 8 outstanding IO requests in the queue. If you recall by grocery check-out metaphor of part 6, this is like having 8 people with full shopping carts waiting in line for a single check-out operator. Since the guy at the back of the line has to wait for the seven people in front of him to be processed, he has to wait longer. So with eight outstanding IO requests, latency should increase as each IO request will be sitting in a queue behind the seven other requests.

By the way, if none of that made sense, then you did not read part 6 and part 7. I urge you to read them before continuing here, because I am assuming prior knowledge of SQLIO and disk latency characteristics and the big trolley theory..

Here is the SQLIO command and below is the result…

SQLIO –kW –b8 –frandom –s120 –t2 –o8 –BH –LS F:\testfile.dat

image

Now take a note of the results reported. IOPS was 526, MBs/sec was 4.11 and as expected, the average latency was much larger than the SQLIO tests we ran in part 7. In this case, latency was 29 milliseconds.

Let’s now compare this to what performance monitor captured. First up, return to Performance Monitor, and stop your data collector set by right clicking on it and choosing “Stop”. Now if you cast your eye to the top left navigation pane, you should see an option called “Reports” listed under “Performance”. Click on “Reports” and look for a sub menu called “User Defined”. Expand “User Defined” and hey presto! Your data collector set should be listed…

image  image

Expand the data collector set and you will find a report for the data you just collected. The naming convention is the server name and the date of the collection. Click on this and you will then see the performance data for that collection in the right pane. At the bottom you can see the six performance counters we chose and just by looking at the graph, you can clearly see when SQLIO started and stopped.

image

Now we have to do one additional step to make sure that we are comparing apples with apples. Performance monitor will calculate its averages based on the total time displayed. As you can see above, I did not run SQLIO straight away, but the performance counters were collected each second nonetheless. Therefore we have a heap of zero values that will bring the averages down and mislead. Fear not though, it is fairly easy (although not completely obvious) to zoom into the time we are interested in. If you look closely, just below the performance graph, where the time is reported, there is a sliding scale. If you click and drag the left and right boxes, you can highlight a specific time you are interested in. This will be shown in the performance graph too, so using this tool, we can get more specific about the time we are interested in. Then in the toolbar above the graph, you will see a zoom button. Click it and watch the magic…

image  image

As you can see below, now we are looking at the performance data for the period when the SQLIO was run. (Now it should be noted that windows performance monitor isn’t particularly granular here. I had to fiddle with the sliding scale a couple of times to accurately set the exact times when SQLIO was started and then stopped.)

image

Now let’s look at the results reported by performance monitor. The screenshot above is looking at the number of Disk Writes per second. Let’s zoom into the figures for the time period and example the average result over the sample period. To save you squinting, I have pasted it below and called out the counter in question. Performance monitor has reported average “Disk Writes/Sec” as 525.423. This is entirely consistent with SQLIO’s reported IOPS of 526.

image

Latency (reported in seconds via the counter Avg. Disk sec/Write) is also fairly consistent with SQLIO. The figure from performance monitor was 0.03 seconds (30 milliseconds). SQLIO reported 29 milliseconds.

image

What about IO size? Well, that’s what Avg disk bytes/write is for… Let’s take a look shall we? Yup.. 8192 kilobytes, which is exactly the parameters specified.

image

SQL IO characteristics revisited (and an awesome script)

Now that we understand what SQLIO is telling us via examining windows performance monitor counters, I’d like to return to the topic of SQL IO patterns. Back at the end of part 6, I spent some time talking about SQL and SharePoint IO characteristics. As a quick recap, I mentioned SQL reads and writes to databases via 8KB pages. Now based on me telling you that, you might assume that if you had to open a large document from SharePoint (say 1MB  or 1024KB), SQL would make 128 IO requests of 8KB each.

While that would be a reasonable assumption, its also wrong. You see, I also mentioned that SQL Server also has a read-ahead algorithm. This algorithm means that means SQL will try and proactively retrieve data pages that are going to be used in the immediate future. As a result, even though a single page is only 8KB, it is not unusual to see SQL read data from disk in a much wider range if it thinks the next few 8KB pages are likely to be asked for anyway. Now as an aside, if you are running SQL Enterprise edition, the possible read-ahead range is from 1 to 128 pages (other editions of SQL max out at 32 pages). Assuming SQL Enterprise edition, this translates to between 8KB and 1024KB for a single IO operation. Think about this for a second… based on the 1MB document example I used in the previous paragraph, it is technically possible that this could be serviced with a single IO request by an enterprise edition of SQL server.

Okay, so essentially SQL has varying IO characteristics when it comes to reading from and writing to databases. But there is still more to it. This is because there are a myriad of SQL IO operations that we did not even consider in part 6. As an example, we have not spoken about the IO characteristics of how SQL writes to transaction logs (which is sequential as opposed to random IO, and does not use 8k pages at all). Another little known fact with transaction logs is that SQL has to wait for them to be “flushed to stable media” before the data page itself can be flushed to stable media. This is known as Write Ahead Logging and is used for data integrity purposes. What is means though is that if logging has a lot of latency, the rest of SQL server can potentially suffer as well (and if it was not obvious before, yet another good reason why people recommend putting SQL data and log files on different disks).

Now I am not going to delve deep into SQL IO patterns any more than this, because we are now getting into serious nerdy territory. However what I will say is this: by understanding the characteristics of these IO patterns, we have the opportunity to change the parameters we pass to SQLIO and more accurately reflect real-world SQL characteristics in our testing. Luckily for all of us, others have already done the hard work in this area. First up, Bob Duffy created a table that summarises SQL Server IO patterns based on the type of operations being performed. Even better than that… Niels Grove-Rasmussen wrote a completely brilliant post, where not only did he list the IO patterns that SQL is likely to exhibit, he wrote a PowerShell script that then runs 5 minute SQLIO simulations for each and every one of them!

I have not pasted the script here, but you will find it at Niels article. What I will say though is that aside from the obvious 8KB random reads and writes that we have concentrated on thus far, Niels listed several other common SQL IO patterns that his SQLIO script tests:

  • 1 KB sequential writes to the log file (small log writes)
  • 64 KB sequential writes to the log file (bulk log writes)
  • 8 KB random reads to the log file (rollbacks)
  • 64 KB sequential writes to the data files (checkpoints, reindex, bulk inserts)
  • 64 KB sequential reads to the data files (read-ahead, reindex, checkdb)
  • 128 KB sequential reads to the data files (read-ahead, reindex, checkdb)
  • 128 KB sequential writes to the data files (bulk inserts, reindex)
  • 256 KB sequential reads to the data files (read-ahead, reindex)
  • 1024 KB sequential reads to the data files (enterprise edition read-ahead)
  • 1 MB sequential reads to the data files (backups)

The script actually handles more combinations than those listed above because it also tests for differing number of threads (-t ) and outstanding requests (-o ). All in all, over 570 combinations of IO patterns are tested. Be warned here… given that each test takes 5 minutes to run by default, with a 60 second wait time in between each test, be prepared to give this script at least 2 days to let it run its course!

The script itself is dead simple to run. Just open a powershell window, and save Niels script to the SQLIO installation folder. From there, change to that directory and issue the command:

./SQLIO_Batch.ps1

Then come back in 3 days! Seriously though, depending on your requirements, you can modify the parameters of the script to reduce the number of scenarios based on editing the first 7 lines of code which is quite self explanatory

$Drive = @('G', 'H', 'I', 'J')
$IO_Kind = @('W', 'R') # Write before read so that there is something to read.
$Threads = @(2, 4, 8)
#$Threads = @(2, 4, 8, 16, 32, 64)
$Seconds = 10*60 # Five minutes
$Factor = @('random', 'sequential')
$Outstanding = @(1, 2, 4, 8, 16, 32, 64, 128)
$BlockSize = @(1, 8, 64, 128, 256, 1024)

Now if this wasn’t cool enough, Niels also written a second script that parses the output from all of the SQLIO tests. This can produce a CSV file that allows you to perform further analysis in excel. To run this script, we need to know the same of the output file of the first script. By default the filename is SQLIO_Result.<date>.txt. For example:

./SQLIO-Parse.ps1 -ResultFileName ‘SQLIO_Result.2010-12-24.txt’

By default the parse script outputs to the screen, but modifying it to write to CSV file is really easy. All one has to do is comment out the second last line of code and uncomment the last one as shown below:

#$Sqlio | Format-Table -Property Kind,Threads,Seconds,Drive,Stripe,Outstanding,Size,IOs,MBs,Latency_min,Latency_avg,Latency_max -AutoSize

$Sqlio | Export-Csv SQLIO_Parse.csv

Below is an example of the report in Excel. Neat eh?

image

Conclusion and coming up next…

By now, you should be a SQLIO guru and have a much better idea of the sort of IO patterns that SQL Server has beyond just reading from and writing to databases. We have covered the IO patterns of transaction logs, as well as examined a terrific PowerShell script that not only runs all of the IO scenarios that you need to, but parses the output to produce a CSV file for deeper analysis. In short, you now have the tools you need to run a pretty good disk infrastructure stress test and start some interesting conversations with your storage gurus.

However at this point I feel there some pieces missing to this disk puzzle:

  1. We have not yet brought the discussion back to lead and lag indicators. So while we know how to hammer disk infrastructure, how can we be more proactive and specify minimum conditions of satisfaction for our disk infrastructure?
  2. Microsoft treatment of disk performance (and in particular IOPS and latency) in their performance documentation is inconsistent and in my opinion, confuses more than it clarifies. So in the next post, we are going to look at these two issues. In doing so, we are going to leave SQLIO and Performance Monitor behind and examine two other utilities including one that is lesser known, but highly powerful.

Until then, thanks for reading

Paul Culmsee

www.hereticsguidebooks.com

 Digg  Facebook  StumbleUpon  Technorati  Deli.cio.us  Slashdot  Twitter  Sphinn  Mixx  Google  DZone 

No Tags

Send to Kindle

Demystifying SharePoint Performance Management Part 7 – Getting at Latency, IOPS and MBPS

This entry is part 7 of 11 in the series Perf
Send to Kindle

Hi all, and welcome to part 7 of this series on SharePoint performance planning. This is the point of the series where I realise that I have much more to write about than I intended. Last time this happened I never got around to finishing the series (*blush* … a certain tribute to a humble leave form ). Like that series, I now have no idea how many posts I will end up doing, but I will keep soldiering on nonetheless.

Recapping the last two posts of this series in particular, we have been looking at the relationship between the performance measures of Disk latency, Disk I/O per second (IOPS) and Disk Megabytes transferred per second (MBPS). We spent most of part 6 looking at the relationship between these three performance metrics by specifically focusing on how the size of an IO request affects things. If you recall, a couple of key points were made:

  • In general, the larger the IO request being made, the more latency there will be, resulting in less IOPS but increased MBPS.
  • Latency is significantly affected by whether an IO requests is sequential or random. To demonstrate this, I used a tool called SQLIO to simulate disk IO to generate some performance stats that demonstrated both IOPS and MBPS improved by some 750% when compared to random IO.

We finished the post by examining the way SQL server performs IO requests and what SharePoint components are IOPS heavy. In short, SQL Server uses a range of request sizes for database reads and writes between 8K and 1024KB. The reason for the range (for reads anyhow) is the read-ahead algorithm (gory detail here), in which SQL attempts to proactively retrieve data that are going to be used in the immediate future. A read-ahead may result in a much larger I/O request being made than a single 8KB page, but much better performance because in effect, SQL is pulling more data from each I/O operation.

In this episode (and the next one)…

Our focus in this post and the next one is similar to part 3, in that we are now going to do some real work and some of it will involve the command line. Therefore also like part 3, if you are one of those project manager types who utilise the wussy “I’m business, not technical” excuse, I want you to persist and try this stuff out. Given that I wrote this series with you in mind, put that damn iPad down, get out your laptop and reload this article! You can try all of the steps below out on your PC while you are reading this.

Now for the tech types reading this, on account of my intention to “demystify” SharePoint performance, I will be more verbose that what you guys need. But consider it this way – I am doing you guys a favour because next time your PM or BA’s eyes start to glaze when you start explaining performance and capacity planning to them, you can point them to this series and tell them that there is no excuse.

This article is going to cover two areas. First up let’s look at what we can do with Windows inbuilt Performance Monitor tool in terms of monitoring Latency and IOPS in particular. Next we will look at a popular tool for stress testing disk infrastructure that gives us visibility into MBPS.

The basics: Performance Monitor 101

Just in case you have never done it before, type in PERFMON on any Windows box at the start button or the command line (by the way, I am assuming Windows 7 or Windows 2008 Server here).

image

If you did that, then you are looking at the classic tool used to understand how a PC or server is performing. Looking at the top left of the resultant window, you should see several options listed under “Performance”. Click on “Performance Monitor” and watch the magic. Congratulations… you now know how to measure CPU as that is the default performance counter displayed.

image

Performance monitor can easily be used to take a look at disk IOPS and latency. Right click on the graph and from the menu choose Add Counters… This will provide you with a long list of “performance objects” (a fancy word for a logical grouping of performance counters)

image

From the list of performance objects, scroll up and find “LogicalDisk”. Move your cursor to the arrow to the right of the LogicalDisk counters and click on it. You should see a list of disk related performance counters appear as shown below.

image   image

Note:  You could have chosen the performance object called PhysicalDisk instead of LogicalDisk. The difference between them is that physical disk counters only consider each hard drive, not the way it is partitioned. The Logical Disk performance object monitors logical partitions of a disk. As a general role (for non techy types reading this), go with LogicalDisk.

Right then… now currently, all of the possible performance counters for LogicalDisk are currently selected, but for now we are only interested in latency and IOPS, which are represented by four counters:

Latency: Avg. Disk sec/Read
Avg. Disk sec/Write
Measures the average time, in seconds, of a read of data to the disk. (Therefore 5ms will be shown as 0.005)
Measures the average time, in seconds, of a write of data to the disk

MS Technet Note: Numbers also vary across different storage configurations (SAN cache size/utilization can impact this greatly)
IOPS Disk Reads/sec:
Disk Writes/sec:
The rate of read operations on the disk per second.
The rate of write operations on the disk per second.
MS Technet Note: This number varies based on the size of I/O’s issued. Practical limit of 100-140/sec per disk spindle, however consult with hardware vendor for more accurate estimation.

Go ahead and select these four counters (use the Ctrl key and click each one to select more than one counter). Now we have to choose which disk or partition that we want to monitor. Below where you chose the performance counters, you will see a label with the suitably unclear title of “Instances of selected object” (I have highlighted it below). From here, choose the hard drive or partition you are interested in. Finally, click the “Add” button at the very bottom and you should see your selected counters listed in the “Added counters” window.

image   image

Click the OK button and you should now be seeing these counters doing their thing. Each performance counter you added is listed below the graph showing the performance data collected in real time. The display shows a time period of 100 seconds and is refreshed each second. Also, a neat feature that some people don’t know about it is to click on one of the counters and then hold down Ctrl and type the letter “H”. This is the shortcut key for highlighting the selected counter and the currently selected counter should now be black. Additionally, you should be able to now use the up and down arrow keys to cycle through the counters and highlight each.

image

At this point, try copying some files or open some applications and watch the effect. You should see a spike in disk related activity reflected in the IOPS and latency counters above. There you go business analysts, you officially have monitored disk performance! Wasn’t so hard was it?

Now that we are monitoring some interesting counters, how about we really give the disk something to chew on! Smile

Upping the ante with SQLIO

SQLIO is an old tool nowadays, but still highly relevant and extremely useful. Despite being named SQLIO, it actually has very little to do with SQL Server! It was provided by Microsoft to help determine the I/O capacity that a server can handle. SQLIO allows you to test a combination of I/O sizes for read/write operations, both sequentially and randomly. Thus, it is useful for stress testing the disk infrastructure for any IO intensive application. Now be warned… you can absolutely smash your disk infrastructure with this tool, so don’t go running this in production without some sort of official clearance. Furthermore, if you want to use SQLIO to test your SAN, be sure to consider the other servers and applications that might be using it. There is potential to adversely affect them.

You can download SQLIO from Microsoft here. It will run on any recent Windows OS, so you can try it on your own PC (now you know why I told you to put your iPad away earlier). Installing SQLIO is very simple, just run SQLIO.MSI and it will install by default into C:\Program Files(x86)\SQLIO folder.

Note: If you want a great tutorial on installing and using SQLIO, look no further than MCM Brent Ozar’s 2009 article entitled SQLIO Tutorial: How to Test Disk Performance).

SQLIO works by reading from and writing to one or more test files, so the first thing we need to do with SQLIO is to set up a configuration file that specifies the location and size of these test files. The configuration file is called PARAM.TXT and is found in the installation folder. Each line of the configuration file represents a test file, its size and a couple of other parameters. The options on each line of the param.txt file are as follows:

  • <Path to test file> Full path and name of the test file to be used.
  • <Number of threads (per test file)>
  • <Mask > Set to 0x0
  • <Size of test file in MB> Ideally, this should be large enough so that the test file will be larger than any cache resident on the SAN (or RAID controller).

Of these 4 parameters, only the first one (the location of the file) and last one (the size of the file) matters for now. Below is a sample param.txt that tests a 20GB file on the E:\ Drive.

image

The next step is to run a  quick SQLIO using sequential writes to create the test file. We are going to use the command-line to do this (although someone has written a GUI for the tool). So open a command prompt, change to the installation directory for SQLIO and type the command below (we will save an detailed explanation of the parameters for later).

sqlio -kW -s10 -fsequential -o8 -b8 -LS -Fparam.txt timeout /T 10

This command will create the file and run a 10 second test. The output will look something like what I have pasted below:

sqlio v1.5.SG

using system counter for latency timings, 2241035 counts per second

parameter file used: param.txt

     file e:\testfile.dat with 1 thread (0) using mask 0x0 (0)

1 thread writing for 10 secs to file e:\testfile.dat

     using 8KB sequential IOs

     enabling multiple I/Os per thread with 8 outstanding

size of file e:\testfile.dat needs to be: 20971520000 bytes

current file size:      104857600 bytes

need to expand by:      20866662400 bytes

expanding e:\testfile.dat …

SQLIO will stop here for a while, while your PC chugs away creating the 20GB test file. Once completed, it will run out quick 10 second test, but you can ignore the rest of the output because this test is  of no consequence.

Running a real test

The previous command was just the entre. We are not interested in the resulting data because the point of the exercise was to create the test file. Now it is time for the main course. Try this command. It will spend 2 minutes running a random IO write to the 20gig test file using a size of 8KB for each write.

sqlio -kW -b8 -frandom -s120 -BH -LS -Fparam.txt

Below is the output that summarises the configuration specified by the above command:

sqlio v1.5.SG

using system counter for latency timings, 2241035 counts per second

1 thread writing for 120 secs to file e:\TestFile.dat

using 8KB random IOs

buffering set to use hardware disk cache (but not file cache)

using current size: 20000 MB for file: e:\TestFile.dat

initialization done

For the next two minutes SQLIO will chug away, hammering the disk with writes. Once the test has been performed, SQLIO will report its findings. You will see IOPS, MBPS and a report of average/max/min latency. On top of this, a histogram showing the distribution of latency is provided. This histogram gives context to the “average latency” figure because it shows the shape of the latency that occurred throughout the test. I graphed the distribution in excel below the SQLIO results below:

CUMULATIVE DATA:

throughput metrics:

IOs/sec:   225.80

MBs/sec:     1.76

latency metrics:

Min_Latency(ms): 0

Avg_Latency(ms): 3

Max_Latency(ms): 111

histogram:

ms: 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24+

%:  4  6  6 31 23 15  5  3  2  1  1  1  1  0  0  0  0  0  0  0  0  0  0  0  0

image

Running the numbers…

Now, before we get into a more detailed test, let’s examine some of the SQLIO parameters:

  • -k specifies whether to perform a read or write test (–kW for write and –kR for read)
  • -s specifies how long to run the test for. In the example above it ran for 2 minutes (120 seconds)
  • -f specifies whether to run a random or sequential IO operation (-frandom)
  • -b specifies the size of the IO operations (in the example above 8KB)
  • -t specifies the number of threads to use. A multi-cpu server should be able to utilise more threads than you have processors. If your storage can handle it, we can increase the number of threads and see what latency arises as a result.
  • -o specifies the number of outstanding requests. This simulates a sudden spike in load and gives an indication of how fast IO requests are being serviced. If you keep adding outstanding requests, latency will start to increase as the number of IO requests outstrips the disks ability to service them.
  • -LS means to capture the disk latency information. If you do not specify this you will not get any latency results

Okay, so how about we see what difference a queue of IO requests makes. Below is a SQLIO command with the addition of the –o parameter. Let’s see what a queue of 4 outstanding requests does and compare the historgram output…

sqlio -kW -b8 -frandom –s120 –o4 -BH -LS -Fparam.txt

And the result? Much more latency than our first example above but no real increase in IOPS or MBPS. Clearly we are already at the limit of what my laptop can handle (I stripped the hyperbole and pasted the counters only).

IOs/sec:   221.73

MBs/sec:     1.73

Min_Latency(ms): 0

Avg_Latency(ms): 17

Max_Latency(ms): 187

 

image

Now since I only changed 1 parameter and such a difference was seen, most people will use SQLIO with a batch file to test different parameters. For example, if you were to paste the commands below into a batch file, you would be running write tests using 16KB, 32KB and 64KB sizes.

sqlio -kW -b16 -frandom -s120 -BH -LS -Fparam.txt

sqlio -kW -b64 -frandom -s120 -BH -LS -Fparam.txt

sqlio -kW -b128 -frandom -s120 -BH -LS -Fparam.txt

For what it’s worth, here is the results for each of the above tests (including the 8KB one we stared with) showing the relationship of IOPS, MBPS and latency. As predicted by our exploration of the relationship between request size, IOPS and MBPS in part 6, latency was smallest with the 8KB option.

8KB write 16KB write 64KB write 128KB write
IOs/sec: 225.80

MBs/sec: 1.76

Avg_Latency(ms): 3

IOs/sec: 220.39

MBs/sec: 3.44

Avg_Latency(ms): 4

IOs/sec: 192.85

MBs/sec: 12.05

Avg_Latency(ms): 4

IO/Sec: 176.30

MB/sec: 22.02

Avg Latency(ms): 5

image

Now one quick note: If you want to play with the –t parameter and add more threads, you will  have to reference the test file directly and not refer to the parameters file. This is because the one of the settings in the param.txt file is the number of threads for each file. Not matter what you put in at the command line, it will be overwritten by what is specified in param.txt. Thus the command below would only run a single thread despite 8 threads being specified via the –t parameter.

sqlio -kW -b64 -frandom -s120 -t8 -o1 -BH -LS -Fparam.txt

sqlio v1.5.SG

using system counter for latency timings, 2241035 counts per second

parameter file used: param.txt

file c:\testfile.dat with 1 thread (0) using mask 0x0 (0)

 

To get around this issue, drop the –F parameter and refer to the test file directly as shown below:

sqlio -kW -b64 -frandom -s120 -t8 -o1 -BH -LS e:\testfile.dat

sqlio v1.5.SG

using system counter for latency timings, 2241035 counts per second

8 threads writing for 120 secs to file e:\testfile.dat

 

Conclusion (and coming up next)…

Phew! Okay, so apart from possibly whetting your appetite for smashing disk infrastructure, you might have also come to the realisation that there are many parameters to test in various combinations. In this entire article, I have assumed random writes to the disk, but what about sequential writes? For that matter, what about reads? What about multiple threads and more outstanding requests? What about longer tests or different sized test files?

These are all important questions and I will answer them… in the next post or two. This one is getting a little too long and I have plenty more to cover in this area.

So have a play with the parameters on SQLIO on your own hardware and in the next post, we will continue looking at SQLIO, plus some great work people have done to make your life much easier using it. I want to also return to PERFMON to show you the relationship between the PhysicalDisk and LogicalDisk counters and what SQLIO reports. Then we will examine two other tools, including one that is lesser known, but a very powerful way to measure disk performance. (That one will redeem me with the tech guys who will have no doubt found this article to be too light on 🙂

Subsequent to that, we hark way back to part 1 and return to a lead indicator point of view of disk IO performance and look at how you can nail the ass off your SAN vendor to ensure they do all the due diligence necessary that your Disk infrastructure will perform well.

Until then, thanks for reading

Paul Culmsee

HGBP_Cover-236x300

www.sevensignma.com.au

 Digg  Facebook  StumbleUpon  Technorati  Deli.cio.us  Slashdot  Twitter  Sphinn  Mixx  Google  DZone 

No Tags

Send to Kindle

Demystifying SharePoint Performance Management Part 6 – The unholy trinity of Latency, IOPS and MBPS

This entry is part 6 of 11 in the series Perf
Send to Kindle

Hi all

Welcome to part 6 on my series in making SharePoint performance management that little more digestible. To recap where we have been, I introduced the series by comparing lead versus lag indicators before launching into an examination of Requests Per Second (RPS) as a performance indicator. I spent 3 posts on RPS and then in the last post, we turned our attention to the notion of latency. We watched a Wiggles Video and then looked at all of the interacting components that work together just to load a SharePoint home page. I spent some time explaining that some forms of latency cannot be reduced because of the laws of physics, but other forms of latency are man made. This is when any one of the interacting components are sub-optimally configured and therefore introduce unnecessary latency into the picture. I then asserted that disk latency was one of the most common area that is ripe for sub-optimal configuration. I then finished that post by looking at how a rotational disk works, the strategies employed to mitigate latency (Cache, RAID, SAN’s etc.)

Now on the note of Cache, RAID and SAN’s Robert Bogue who I mentioned in part 1, has also just published an article on this topic area called Computer Hard Disk Performance – From the Ground Up. You should consider Robert’s article part 5.5 of this series of posts because it expands on what I introduced in the last post and also spans a couple of the things I want to talk about in this one (and goes beyond it too). It is an excellent coverage of many aspects of disk latency and I highly recommend you check it out).

Right! In this post, where will look more closely at latency and understand its relationship with two other commonly cited disk performance measures: IOPS and MBPS. To do so, lets go shopping!

Why groceries help to explain disk performance

image

Most people dislike having to wait in a line for a check-out at a supermarket and supermarkets know this. So they always try and balance the number of open check-out counters so that they can scale when things are busy, but not pay the operators to standing around when its quiet. Accordingly, it is common to walk into a store when its quiet and only find only one or two check-out counter open, even if the supermarket has a dozen or more of them.

The trend in Australian supermarkets nowadays is to have some modified check-out counters that are labelled as “express.” In these check-outs, you can only use them if you are buying 15 items or less. While the notion of express check-outs has been around forever, the more recent trend is to modify the design of express check-out counters to have very limited counter space and no moving roller that pushes your goods toward the operator. This discourages people with a fully-loaded trolley/cart to use the express lane because there is simply not enough room to unload the goods, have them scanned and put them back in the trolley. Therefore, many more shoppers can go through express counters than regular counters because they all have smaller loads.

This in turn frees up the “regular” check-out counters for shoppers with a large amount of goods. Not only do they have a nice long conveyor belt with plenty of room for shoppers to unload all of their goods onto and rolls to the operator, but often there will be another operator who puts the goods into bags for you as well. Essentially this counter is optimised for people who have a lot of goods.

Now if you were to measure the “performance” of express lanes versus regular lanes, I bet you would see two trends.

  • Express lanes would have more shoppers go through them per hour, but less goods overall
  • Regular lanes would have more goods go through them per hour, but less shoppers overall

With that in mind, lets now delve back into the world of disk IO and see if the trend holds true there as well.

Disk latency and IOPS

In the last post, I specifically focused on disk latency by pointing out that most of the latency in a rotational hard drive is from rotation time and seek time. Rotation time is time taken for the drive to rotate the disk platter to the data being requested and seek time is how long it takes for the hard drive’s read/write head to then be positioned over that data. Depending on how far the rotation and head have to move, latency can vary. Closely related to disk latency is the notion of I/O per second or “IOPS”. IOPS refer to the maximum number of reads and writes that can be performed on a disk in any given second. If we think about our supermarket metaphor, IOPS is equivalent to the number of shoppers that go through a check-out.

The math behind IOPS and how latency affects it is relatively straightforward. Let’s assume a fixed latency for each IO operation for a moment. If for example, your disk has a large latency… say 25 milliseconds between each IO operation, then you would roughly have 40 IOPS. This is because 1 second = 1000 milliseconds. Divide 1000 by 25 and you get 40. Conversely, if you have 5 milliseconds latency, you would get 200 IOPS (1000 / 5 = 200).

Now if you want to see a more detailed examination of IOPS/ latency and the maths behind it, take a look at an excellent post by Ian Atkin. Below I have listed the disk latency and IOPS figures he posted for different speed disks. Note that a 15k RPM disk came in at around 175-210 IOPS which suggests a typical latency average of between 4.7 and 5.7 milliseconds. (1000/175 = 5.7 and 1000/210 = 4.7). Note: Ian’s article explains in depth the maths behind the average calculation in this section of his post.

image

The big trolley theory of IOPS…

While that math is convenient, the real world is always different to the theoretical reality I painted above. In the world of shopping, imagine if someone with one or two trolleys full of goods like the picture below, decided to use the express check-out. It would mean that all of the other shoppers have to get annoyed and wait around for this shoppers goods to be scanned, bagged and put back into trolley. The net result of this is a reduced number of shoppers going through the check-out too.

image

While the inefficiencies of a supermarket is something that is easy to visualise for most people, disk infrastructure is less so. So while the size of our trolley has an impact on how many people come through a check-out, in the disk world, the size of the IO request has precisely the same effect. To demonstrate, I ran a basic test using a utility called SQLIO (which I will properly introduce you to in part 7) on one of my virtual machines. Below is the results of writing data randomly to a 500GB disk. In the first test we wrote to the disk using 64KB writes and in the second test we used 4KB writes. The results are below:

Size of Write IOPS Result
64KB 279
4KB 572

Clearly, writing 4KB of data over time resulted in a much higher IOPS than when using 64KB of data. But just because there is a higher IOPS for the 4KB write, do you think that is better performance?

Disk latency and MBPS

So far the discussion has been very IOPS focussed. It is now time to rectify this. In terms of the SQLIO test I performed above, there was one other performance result I omitted to show you – the Megabytes per second (MBPS) of each test. I will now add it to the table below:

Size of Write IOPS Result MBPS Result
64KB 279 17.5
4KB 572 2.25

Interesting eh? This additional performance metric paints a completely different picture. In terms of actual data transferred, the 4KB option did only 2.25 megabytes per second whereas the 64KB transferred almost 8 times that amount! Thus, if you were judging performance based on how much data has been transferred, then the 4KB option has been an epic fail. Imagine the response of 500 SharePoint users, loading the latest 30 megabyte annual report from a document library if SharePoint used 4KB reads … Ouch!

So the obvious question is why did a high IOPS equate to a low MBPS?

The answer is latency again (yup – it always comes back to latency). From the time the disk was given the request to the time it completed, writing 4KB simply doesn’t take as long to write as 64KB does. Therefore there are more IOPS that take place with smaller writes. Add to that, the latency from disk rotation and seek time per IO operation and you start to see why there is such a difference. Eric Slack at Storage Switzerland explains with this simple example:

As an illustration, let’s look at two ways a storage system can handle 7.5GB of data. The first is an application that requires reading ten 750MB files, which may take 100 seconds, meaning the transfer rate is 75MB/s and consumes 10 IOPS. The second application requires reading ten thousand 750KB byte files, the same amount of data, but consumes 10,000 IOPS. Given the fact that a typical disk drive provides less than 200 IOPS, the reads from the second application probably won’t get done in the same 100 seconds that the first application did. This is an example of how different ‘workloads’ can require significantly different performance, while using the same capacity of storage.

Now at this point if I haven’t completely lost you, it should become clear that each of the unholy trinity of latency, IOPS and MBPS should not be judged alone. For example, reporting on IOPS without having some idea of the nature of the IO could seriously mislead. To show you just how much, consider the next example…

Sequential vs. Random IO

Now while we are talking about the IO characteristics of applications, two really important point that I have neglected to mention so far is the range of latency and the impact of sequential IO.

The latency math I did above was deliberately simplified. Seek and rotation time are actually across a range of values because sometimes the disk does not have to rotate the spindle/move the head far. The result is a much reduced seek latency and accordingly, increased IOPS and MPBS. Nevertheless, the IO is still considered random.

Taking that one step further, often we are dealing with large sections of contiguous space on the hard disk. Therefore latency is reduced further because there is virtually no seek time involved. This is known as sequential access. Just to show you how much of a difference sequential access makes, I re-ran the two tests above, but this time writing to sequential areas of the disk and not random. With the reduced seek and rotation time, the difference in IOPS and MBPS is significant.

Size of Write IOPS Result MBPS Result
64KB 2095 131
4KB 4152 16

The IOPS and subsequent MBPS has improved significantly from the previous test to the tune of a 750% improvement. Nevertheless, the size of the request and its relation to IOPS and MPBS still holds true. The smaller the size of the IO request being read or written, the more IOPS requests can be sustained, but the less MBPS throughput can be achieved. The reverse then holds true with larger IO requests.

One conclusion that we can draw from this is that specifying IOPS or MBPS alone has the potential to really distort reality if one does not understand the nature of the IO request in terms of its characteristics. For example: Let’s say that you are told your disk infrastructure has to support 5000 IOPS. If you assumed a 4K IO size that is accessed sequentially, then far fewer disks would be required to achieve the result compared to a 64KB IO accessed randomly. In the 64KB case, you would need many disks in an array configuration.

SQL IO Characteristics

So now we get to the million dollar question. What sort of IO characteristics does SQL and SharePoint have?

I will answer this by again quoting from Ian Atkin’s brilliant “Getting the Hang of IOPS” article. Ian makes a really important point that is relevant to SQL and SharePoint in his article which I quote below:

The problem with databases is that database I/O is unlikely to be sequential in nature. One query could ask for some data at the top of a table, and the next query could request data from 100,000 rows down. In fact, consecutive queries might even be for different databases. If we were to look at the disk level whilst such queries are in action, what we’d see is the head zipping back and forth like mad -apparently moving at random as it tries to read and write data in response to the incoming I/O requests.

In the database scenario, the time it takes for each small I/O request to be serviced is dominated by the time it takes the disk heads to travel to the target location and pick up the data. That is to say, the disk’s response time will now dominate our performance.

Okay, so we know that SQL IO is likely to be random in nature. But what about the typical IO size?

Part of the answer to this question can be found in an appropriately titled article called Understanding Pages and Extents. It is appropriate because as far as SQL server database files and indexes are concerned, the fundamental unit of data storage in SQL Server is an 8KB page. The important point for our discussion is that Disk I/O many read and write operations are performed at the page level. Thus, one might assume that 8KB should be the size assumed when working with IOPS calculations because it is possible for SQL to write 8KB to disk at a time.

Unfortunately though, this is not quite correct for a number of reasons. Firstly, eight contiguous 8KB pages are grouped into something called an extent. Given than an extent is a set of 8 pages, the size of an extent is 64KB. SQL Server generally allocates space in a database on a per-extent basis and performs many reads across extents (64KB). Secondly, SQL Server also has a read-ahead algorithm that means SQL will try and proactively retrieve data pages that are going to be used in the immediate future. A read-ahead is typically from 1 to 128 pages for most editions which translates to between 8KB and 1024KB. (for the record, there is a huge amount of conflicting information online about SQL IO characteristics. Bob Door’s highly regarded SQL Server 2000 I/O basics article is the place to go for more gory detail if you find this stuff interesting).

A read-ahead interlude…

Before we get into SharePoint disk characteristics, it is worthwhile mentioning a great article by Linchi Shea called Performance Impact: Some Data Points on Read-Ahead.  Linchni did an experiment by disabling read-ahead behaviour in SQL Server and measured the performance of a query on 2 million rows. With read-ahead enabled, it took 80 seconds to complete. Without read-ahead it took 210 seconds. The key difference was the size of the IO requests. Without read-ahead the reads were all 8KB as per page size. With read-ahead, it was over 350KB per read. Linchi makes this conclusion:

Clearly, with read-ahead, SQL Server was able to take advantage of large sized I/Os (e.g. ~350KB per read). Large-sized I/Os are generally much more efficient than smaller-sized I/Os, especially when you actually need all the data read from the storage as was the case with the test query. From the table above, it’s evident that the read throughput was significantly higher when read-ahead was enabled than it was when read-ahead was disabled. In other words, without read-ahead, SQL Server was not pushing the storage I/O subsystem hard enough, contributing to a significantly longer query elapsed time.

So for our purposes, lets accept that there will be a range of IO sizes for read/writes to databases between 8KB to 1024KB. For disk IO performance testing purposes, lets assume that much of this is across the extent boundaries of 64KB. Based on our discussion of latency and MBPS where the larger the IO being worked with, the lower the IOPS, we can now get a better sense of just how much disk might need to be put into an array to achieve a particular IOPS target. As we saw with the examples earlier in this post, 64KB IO sizes result in more latency and lower IOPS. Therefore SharePoint components requiring a lot of IOPS may need some pretty serious disk infrastructure.

SharePoint IO Characteristics

This brings us onto our final point for this post. We need to understand what SharePoint components are IO intensive. The best place to start to determine this is page 29 of Microsoft’s capacity planning guide as it supplies a table listing the general performance requirements of SharePoint components. A similar table exists on page 217 of the Planning guide for server farms and environments for Microsoft SharePoint Server 2010. We will finish this post with a modified table that shows all the SharePoint components listed with medium to high IOPS requirements from the capacity planning guide, along with some of the comments from the server farm planning guide. This gives us some direction as to the SharePoint components that should be given particular focus in any sort of planning. Unfortunately, IOPS requirements are inconsistently written about in both documents. Sad smile

Service Application

Service Description

SQL Server IOPS

SharePoint Foundation Service

The core SharePoint service for content collaboration.

Almost all of the IOPS occurs in SharePoint content databases. IOPS requirements for content databases vary significantly based on how your environment is being used, and how much disk space and how many servers you have. Microsoft recommends that you compare the predicted workload in your environment to one of the solutions that they have tested. I will be covering this in part 8.

XXX

Logging Service

The service that records usage and health indicators for monitoring purposes.

The Usage database can grow very quickly and require significant IOPS. Use one of the following formulas to estimate the amount of IOPS required:
115 × page hits/second
5 × HTTP requests

XXX

SharePoint Search Service

The shared service application that provides indexing and querying capabilities. There is a dedicated document that among other things that covers IOPS requirements.

For the Crawl database, search requires from 3,500 to 7,000 IOPS.
For the Property database, search requires 2,000 IOPS.

XXX

User Profile Service

The service that powers the social scenarios in SharePoint Server 2010 and enables My Sites, Tagging, Notes, Profile sync with directories and other social capabilities

No mention of IOPS is made in both the planning guides

XXX

Web Analytics Service

The service that aggregates and stores statistics on the usage characteristics of the farm.

The planning guide suggests readers consult a dedicated planning guide for web analytics, but unfortunately no mention of IOPS is made, let alone a recommendation 

XXX

Project Server Service

The service that enables all the Microsoft Project Server 2010 planning and tracking capabilities in addition to SharePoint Server 2010

No mention of IOPS is made in both the planning guides

XXX

PowerPivot Service

The service to display PowerPivot enabled Excel worksheets directly from the browser

No mention of IOPS is made in both the planning guides

XX

(In case it is not obvious, XX – Indicates medium IOPS cost on the resource and XXX indicates high IOPS cost on the resource)

Conclusion (and coming up next)

Whew! I have to say, that was a fairly big post, but I think we have broken the back of latency, IOPS and MBPS. In the next post, we will put all of this theory to the test by looking at the performance counters that allow us to measure it all, as well as play with a couple of very useful utilities that allow us to simulate different scenarios. Subsequent to that, we will look at these measures from a lead indicator perspective and then examine some of Microsoft’s results from their testing.

Until then, thanks very for reading. As always, comments are greatly appreciated.

Paul Culmsee

www.hereticsguidebooks.com

 Digg  Facebook  StumbleUpon  Technorati  Deli.cio.us  Slashdot  Twitter  Sphinn  Mixx  Google  DZone 

No Tags

Send to Kindle

Demystifying SharePoint Performance Management Part 5 – So what is latency anyway?

This entry is part 5 of 11 in the series Perf
Send to Kindle

Hi all

Welcome to part 5 in my attempt to make SharePoint performance management a little more accessible. Now that we have dealt with the world of request per second in parts two, three and four, we will focus our attention somewhere different for a post or three.

To set the scene, we are going to take a bit of an end to end look at what it takes to load a SharePoint page. I suspect some readers do not have the full picture on just how many components interact together just to load the SharePoint home page. Things are much more complex in reality than the typical architectural view that adorns SharePoint blogs. A typical SharePoint diagram will list the servers and their roles, but what about all…

  • the network gear like routers, switches, reverse proxies and firewalls that are part of the mix?
  • the VMWare or HyperV virtual hosts that provide the virtualised servers? And
  • the storage area network and its associated paraphernalia that these virtual servers make use of for disk infrastructure?

Make no mistake people, configurations these days are hugely complex and have many moving parts. If any of the various components listed above were to fail or become a bottleneck, the performance of the entire system suffers. Therefore, we need assurance that each component has been optimised to ensure overall function.

This brings us onto the topic of latency. If you are not sure what latency is, I can guarantee that you actually do know. You see, if you have ever experienced a jittery skype call, or your pornography is slow to load, or you have watched a roving reporter respond several seconds after being asked a question from the studio, you are experiencing latency.

Now, the important point to make straight up is that latency is unavoidable because of the laws of physics. Take the example of one of the rovers that NASA sent to Mars. All radio signals to Mars travel at the speed of light (which despite Star Trek’s best efforts to persuade us otherwise, is the absolute speed limit of the universe). The speed of light is around 300,000 kilometres per second and the distance to Mars is currently around 150 million kilometres from Earth. So doing some basic math, we find that it takes a little over 8 minutes for a signal to get from Earth to Mars.

  • 150,000,000 / 300,000 = 500 seconds
  • 500 / 60 = 8.3 minutes

In this example of latency, no matter what  happens, there will always be around 8 minutes of latency between the time an instruction is sent to a rover, to the time it receives and acts on it. Unless Einstein was wrong, this isn’t about to change in a hurry either.

A “lean” view of latency…

Latency is a concept that extends beyond the forces of nature. Let me give you another form of latency that I am sure you have experienced, using Microsoft as the straw man. Let’s say you have a problem with SharePoint and you log a call with Microsoft or your support provider. You call the technical support line and after twiddling your thumbs in the telephone queue for an eternity, you get an inexperienced level 1 tech, who doesn’t understand your problem at all and is hell bent on closing your call anyway because someone higher up in the organisation actually believed that call-time is an indicator of happy customers. You repeat yourself each and every time as your call is slowly routed up the tech support hierarchy. Finally, by the time you get to level 3 or 4, you finally get a good tech who gives you the quick answer you were looking for. The problem is that three weeks have passed to get to this point.

This is also a form of latency. But unlike the first example. It was not the laws of nature this time, but man made laws that caused wasted time. I will call it organisational latency. Addressing this form of latency is a multi billion dollar industry, and keeps an army of organisational/process improvement consultants busy, trying to reduce wastage and improve customer outcomes (now you know what Lean is all about if you hear people taking about it).

So, returning to the SharePoint context – we have a lot of moving parts. We know we cannot alter the laws of physics, but how do we know whether all of the various components are working to their optimum level? Is there any man-made latency that we could reduce or eliminate?

Oh, yes, indeedie there is… and to put some context  to it, let’s utilise the musical genius that is the Wiggles. I found that their rendition of the old folk song “Dem bones” serves my purpose nicely.

 

The Wiggles, teaching us about SharePoint latency 🙂

When you perform the seemingly benign task of requesting a page with your browser, an amazing number of things have to happen. The browser forms a HTTP request and then passes this to the TCPIP stack on your PC, which takes the HTTP request and breaks it up into IP packets. These packets are passed to your network card driver that turns these packets into Ethernet frames and sends them over the wire. Each network device (switch, router, etc.) has to process each frame or IP packet and to work out where to forward it. Eventually the request finds it way to the destination server where the Ethernet frames are stripped, the IP packets are reassembled into the original HTTP request, passed to IIS and SharePoint then acts on it.

Now all I described above was the task of getting a HTTP request from a browser to a server. To see the entire picture, let’s all sing along with the Wiggles shall we? We will assume a two server deployment, utilising a VMware based virtual web front end SharePoint server and a physical SQL Server. Both servers use a Storage Area Network (SAN) for disk. Cue the melody from “Dem Bones”…

  • Your PC connects to a distribution switch
  • The distribution switch is connected to the core switch
  • The core switch connects to the HyperV host
  • The HyperV host connects to the virtual Web Front End Virtual Machine

… okay so we have managed to get from our browser to the SharePoint web front end but at this point, the web front end hasn’t really done anything yet.  In terms of latency, we had to get through the switches, as well as the virtualisation infrastructure to the virtual SharePoint web front end box. The switches had very little latency at all – probably around 1-2 microseconds (which translates about 0.001 to 0.002 milliseconds) for a network packet to go in one port and out the other. The virtualisation infrastructure also introduced some latency, because there is overhead in running a virtual machine within a physical machine. However, assuming it is well configured and that there aren’t too many virtual machines competing for physical resources like CPU and memory, then that latency is fairly negligible.

Now the virtual web front end server needs to actually deal with the request from your PC. This involves pulling data from the disk infrastructure, so back to the Wiggles we go…

  • the Web Front End Virtual Machine connects to the HyperV host
  • The HyperV host connects to the SAN Switch
  • The SAN Switch connects to the Storage Array
  • The Storage Array connects to the Web Front End disk
  • The Web Front End disk returns data to the SAN Switch
  • The SAN switch returns data to the HyperV host
  • The HyperV host returns data to the Web Front End Virtual Machine

…at this point, the web front end server has retrieved any data it needs to from the disk subsystem. There was definitely latency here. The SAN switches have a similar latency to network switches which is negligible, but the physical disks on the SAN are another story (which we will get to soon). But wait a second – that just loads the stuff the web front end server stores or caches locally, as well as writing to the IIS and SharePoint logs. What about all those sexy web parts you have on the front page that aggregate the latest news feed? That stuff needs to pull information from the SharePoint content database on the SQL Server. So let’s continue, now incorporating the connection between the virtual web front end and SQL Server (Remember, I am assuming the SQL box is not virtualised).

  • The Web Front End Virtual Machine connects to SQL box (via the network on TCPIP port 1433)
  • The SQL Box connects to the SAN Switch
  • The San Switch connects to the Storage Array
  • The Storage Array connects to the SQL disk
  • The SQL disk returns data to the SAN Switch
  • The SAN switch connects to the SQL box
  • The SQL Box connects to Web Front End Virtual Machine (via the network on TCP port 1433)
  • The Web Front End Server returns the page to your PC (via the network on TCP port 80)

Now at this point, non tech oriented readers might be thinking, “Bloody hell! I didn’t realise there were that many interactions.” For you guys… now you know why tech guys are the way they are. Tech guys reading this would know full well that I glossed over many things. For example, I did not include the authentication process in the sequence above, nor did I describe important virtualisation aspects such as VM memory compression. On top of that I glossed big-time over the full gamut of SAN I/O paths.

There is a form of man-made latency that can occur in any of these steps outlined above as a result of the complexity. It is very easy to overlook an important aspect, or to misconfigure something or to assume the default configuration is optimal. In my consulting experience I have seen sub optimal configuration in all of the above touchpoints, but out of all of them, there is one area that is far more likely to have latency issues than any of the other areas: The disk infrastructure.

We will round out this post by taking a fairly high level view at disk infrastructure and why it is latency prone.

Understanding disk latency

Below is a Wikipedia picture that shows the essential components of most hard drives. This type of hard drive is really not that different from its original design in 1954. It is called a rotational hard drive and the spindle rotates, while the actuator arm moves the head to the right position to read data off the platter. As you can imagine, this happens pretty fast too. Most high end hard drives spin the platter at 15000RPM – dizzying, eh?

 

But to put disk performance in perspective, consider my previous example of a network switch with a 1-2 microsecond latency to process an Ethernet frame as it transits through one network port to another. By comparison, a modern hard drive takes a hell of a lot longer to do what it needs to do. As a simple example, the time taken just for the drive to rotate the disk platter takes around 2 milliseconds (or 2000 microseconds). Not only is this a staggering 2000 times slower than the network switch but it does not take into account the time it takes for the hard drive’s read/write head to then be positioned over the sector (this is called seek time and can take anywhere between 3 and 15 milliseconds).

This latency clearly is problematic, and vendors compensate by utilising multiple sets of disks and liberal use of cache technology to mitigate it. Imagine putting 10 hard disks together and when data is saved, parts of it is written to each hard disk. Now you have reduced latency because each drive is handling a smaller part instead of a single drive handling it all. It is important to note that we have done nothing about laws of physics latency per single drive (thanks Robert Bogue for pointing that out) , but throughput induced latency has reduced by using them all. It is just like when you are out the supermarket and there are ten check-out operators working instead of 1. The wait times are much shorter because there are more check out operators available to service the request. (This is the essence of RAID technology and should be familiar to most readers).

But there is still more to the latency story than disks taking time to do their thing. At the operating system level, there are various layers and drivers doing stuff. I won’t go too much into this is except to suggest that if the world of the Class Drivers, Port Drivers, Device Miniport Drivers and Disk Subsystems rock your world then Jeff Hughes has a great writeup where he describes the whole Windows disk IO system in detail.

A note on SSD

I would be remiss not to make a point about these newfangled Solid State Drives (You might have heard them mentioned as SSD). This is a newer technology for hard drives that do not employ any moving mechanical components, like platters and movable read/write heads. Solid State Drives have some seriously improved performance in terms of latency, because they store the data in persistent memory. Wikipedia cites that SSD latency is around 0.1 millisecond compared to rotational drives being around 5-10 milliseconds. The downside is that they are more expensive than traditional rotational disks. According to a May 2012 article, SSDs cost approximately US$0.65 per GB whereas traditional hard disks cost about US$0.05 per GB. Expect prices to continue to fall and for them to appear in more and more solutions.

Then there are SANs

In terms of disk infrastructure and latency aspects, most organisation’s utilise a Storage Area Network (SAN) topology. I previously mentioned the idea of RAID configurations that make use of multiple disks to improve latency (among other things). SANs take the RAID idea and abstracts it further as shown below.

image

(credit for this image is Orbis solutions: http://orbissolutionsinc.com/blog/tag/storage-arrays/)

I sometimes describe SANs to people as a “fridge full of hard drives connected to multiple servers”. What the above diagram shows is that the disks are physically not connected to the servers that use them. Instead they are connected to a storage array via cables, with a switch or three in between. Each server has some disk space reserved for it on the SAN. So the result is we have one centralised high performing disk array where we can take advantage of all of the disks housed within.

But it’s important to understand here that each time data is read from or written to disk, it passes across those cables and through the switches. Like an internet connection, the SAN switch and cables not only have bandwidth limitations, but are prone to misconfiguration. Imagine 50 servers writing data at the same time. If things are not well configured, the SAN switch infrastructure might become overwhelmed like a freeway during peak hour. Direct attached storage (i.e. – the hard drive or RAID array is plugged into the server directly) typically have a higher bandwidth. This quote from a nice sqlteam.com article on SAN performance explains it well.

For instance, if a server is equipped with two older 1-Gbps host bus adapters (HBAs), its MBps throughput would be capped at about 200MB per second no matter how powerful the rest of the SAN is. Replacing the 1-Gbps HBAs with two newer 4-Gbps HBAs or adding more HBAs may improve the throughput, if the HBAs are indeed the throughput bottleneck. But the SAN drive throughput could also be limited by the maximum throughput of the inter-switch links in the SAN switched fabric. Further down the I/O paths, the front-side adapter ports on the disk array, the cache in the disk array, the disk controllers, and the disk spindles can all become the bottleneck limiting the MBps throughput of the SAN drive.

Conclusion and coming next…

Okay… at this point let’s take a breather. For the tech guys reading this post, none of what I covered may seem particularly earth shattering, but it was important to set the context for a deeper dive into disk latency in the next couple of posts. If you are not normally of the tech persuasion, then I hope that this post has opened your eyes a little to just how complicated deployments can be and accordingly, how hard it can sometimes be to pinpoint latency issues when they occur.

In the next post, we will take a deeper look at disk latency and its relationship to the indicators of IOPS and MBPS. We will then examine tools to measure latency and how to best use it as a lead indicator.

Until then, thanks for reading and be sure to check out my recent business book “The Heretics Guide to Best Practices

Paul Culmsee

www.sevensigma.com.au

 Digg  Facebook  StumbleUpon  Technorati  Deli.cio.us  Slashdot  Twitter  Sphinn  Mixx  Google  DZone 

No Tags

Send to Kindle