Tuesday, May 27, 2014

Cloudy I/O Performance - Increasing Azure IOPS (Part 2 of 2)

Notes: This is part 2 of a 2 part post. You can find part 1 here. Since publication of this article Microsoft has introduced new machine types with higher performing storage. More information can be found here.


In the last article we discussed a repeatable testing methodology to quantify storage performance in the cloud, and in this article we'll put that methodology into practice. I've done substantial testing in Azure and aim to illustrate what your options are for scaling performance at this point in time.


I undertook this project to see what can be done to increase disk I/O in Windows Azure IaaS. Upon researching the topic I found several interesting articles. Among those are:
There seems to be little consensus regarding disk striping in Windows Azure IaaS. Some blogs recommend this while some of Microsoft's own writing seems to discourage it. After combing through the options the following points stand out:
  • Disk Striping (Software RAID 0) may or may not increase performance based on your workload.
  • Striping will increase I/O capacity to a degree (which we'll test here).
  • What software striping solution works better: legacy (Windows software RAID from 2000 to present) or Storage Spaces (new software "RAID" in Windows 2012 and up)?
  • How does NTFS cluster size impact performance?
  • If striping, disable geo-replication as Microsoft explicitly warns against the use of geo-replication with this solution.
  • If possible, use native application load distribution rather that disk striping to split I/O.  (For example, split DB files in SQL across disks)
  • Some articles reference needing to use multiple storage accounts to get maximum performance. This is not true; as of 6/7/2012 storage account targets are 20,000 IOPS per account. Unless you will exceed the 20,000 keep all your disks on one account for the sake of simplicity. We will prove that does not have an impact on performance.
With that said, I want to quantify the solution for my given scaling problem with the notion that if the tests are simple enough to run, this approach can be used for any future scaling problem as well.

Putting it All Together

We'll use the testing methodology outlined in part 1 of this article to collect the results. In this case we need to first add disks and set up stripes in Azure Windows VMs.
Note: To jump straight to Azure disk performance tests, scroll to the bottom of this article.

Create New Disks and Attach to Designated VM

In order to run all the tests listed below, you need to know how to create new disks and attach them to your virtual machine. My favorite solution to this is to use a locally created dynamic VHD and upload it to the location you would like using PowerShell. Let's go through the process of attaching one disk as a primer:
  1. Decide which storage account you will use for these disks. If you plan on doing striping of any kind, ensure the storage account is set to "Locally Redundant" replication (Storage->Desired Storage Account->Configure), as "Geo Redundant" is not supported. Since the replication setting applies to all blobs (Azure's terminology; disks) in that account you may want to have a dedicated account for these disks to keep your others Geo Redundant.

  2. Determine what container you would like to store your Azure disk blob by opening the Azure management portal and navigating to Storage->Desired Storage Account->Containers and copy the URL to your clipboard. To keep things simple you may want to create a new storage container, so do so now and use that URL if desired.
  3. Using Hyper-V (On Windows 2008 or higher including Windows 8) create an empty dynamically expanding VHD disk of your desired size. For my testing I have been using 10GB disks. Note 1: Do not create a VHDX; Azure uses the older VHD format. Note 2: You'll need to re-create the VHD for each disk if you intend on using Storage Spaces as each disk must have a unique ID. 
  4. #create a dynamically expanding 10GB VHD; change size as appropriate
    New-VHD –Path $sourceVHD –SizeBytes 10GB -Dynamic
  5. This disk will be uploaded to the container we selected in step 1. Determine the name you want the disk to be referenced by in Azure and execute the following script:
  6. #import Azure cmdlets
    import-module azure.psd1
    #specify your subscription so PS knows what account to upload the data to
    select-azuresubscription "mysubscriptionname"
    #$sourceVHD should be the location of your empty vhd file
    $sourceVHD = "D:\Skydrive\Projects\Azure\AzureEmpty10G_Disk.vhd"
    #$destinationVHD should be the URL of the container and the name of the vhd you want created in your account. Obviously for subsequent disks you need to change the VHD name. 
    $destinationVHD = "http://yourstorageacctname.blob.core.windows.net/vhds/data02.vhd"
    #now upload it. 
    Add-AzureVhd -LocalFilePath $sourceVHD -Destination $destinationVHD

  7. Add this new disk as available to VMs by navigating to Virtual Machines->Disks->Create

  8. Enter the desired management name for this disk and input or browse to the URL of the VHD you just uploaded and click the check box.
  9. Attach the disk to your VM by navigating to Virtual Machines->Ensure your desired VM is highlighted->Attach->Attach Disk

  10. Select the disk we just added. Your cache preference will depend on the application. In my case this is off but you will want to use the methodology outlined in the first part of this article to test caching impact for your application. Note a change of cache status requires a VM reboot.
Now for a brief tutorial on how to set up our two types of striped disks; you'll likely only be using one of the two but I'll cover both just in case. Performance results of each are outlined later in this article.

Set Up a Traditional Software Stripe in Windows

Setting up a traditional software stripe is easy. I've tested this on Windows 2003 and higher.

  1. Logon to your VM as an admin and open the Disk Management tool.
  2. If prompted, allow the initialization of the disks.
  3. Right-click on one of the newly created empty volumes and select New Striped Volume.

  4. Select the desired disks and continue.

  5. Create and format a new NTFS disk using your striped volume. Make sure to pay attention to the cluster size (results below).

Setup a Storage Spaces Software Stripe in Windows 2012 or Higher

Microsoft introduced a new approach to disk pooling in Windows Server 2012 and Windows 8 called Storage Spaces. This interesting new tech allows for a myriad of different configuration options including disk tiering which can be useful for on-premise servers. In this case we'll be using the "simple" pool type which is similar to disk striping.
  1. Open Server Manager and navigate to File and Storage Services -> Volumes -> Storage Pools
  2. Under Storage Pools you should see "Primordial". (As opposed to "Unused Disks". I'm guessing someone was pretty proud of that.) Right click it and select "New Storage Pool".

  3. Walk through the Wizard selecting each disk you would like to be part of the pool.

  4. On the results page, ensure "Create a virtual disk when this wizard closes" is selected and click "Close".

  5. Walk through the Virtual Disk Wizard, specifying a meaningful name and selecting simple storage layout and fixed provisioning.

  6. On the results page, ensure "Create a volume when this wizard closes" is selected and click "Close".
  7. Complete the New Volume Wizard specifying your desired drive letter and desired NTFS cluster size.

Run Tests/Collect Results!

Now that we have our disks configured, we need to run our tests. For instructions how how to do so, see part 1 of this topic here.
When analyzing the IOMeter output you will want to pay special attention to the following metrics:
  • IOPS (Cumlative, Read, Write, Higher is better)
  • MBps (Cumlative, Read, Write, Higher is better)
  • Response Time (Avg, Avg Read, Avg Write, lower is better) 
If putting the data together for a report, Excel works nicely as I'll display below.


Now for the most important part, the findings. Tests performed:

Sector Size Tests:

  • 1 Disk, 4k Sector Size (default)
  • 1 Disk, 8k Sector Size
  • 1 Disk, 16k Sector Size
  • 1 Disk, 32k Sector Size
  • 1 Disk, 64k Sector Size
  • 3 Disks, 4k Sector Size (results confirmation test)
  • 3 Disks, 32k Sector Size (results confirmation test)

Table 1-Cluster Size Tests
Table 2-Cluster Size Verification
Sector size tests echo what others have observed with Azure; since IOPS are capped at 500 (or 300 for basic VMs) larger sector sizes can result in higher throughput. In my case 32k was the sweet spot; depending on your workload your results will vary slightly. I have seen consistently (albeit slightly) higher performance with larger sector sizes in Azure.

Legacy Disk Striping Tests:

  • 1 Disk, 32k Sector Size
  • 2 Disks, Striped Volume, 32k Sector Size
  • 2 Disks in 2 Storage Accounts, Striped Volume, 32k Sector Size (Multiple Storage Account Test)
  • 3 Disks, Striped Volume, 32k Sector Size
  • 4 Disks, Striped Volume, 32k Sector Size
<See Bar Charts Below Under Disk Striping Methodology>
Table 3-Legacy Striping and Storage Account Tests
You can see with one disk we get 500 IOPS as expected. From there we can see a scaling trend that is most definitely not linear. Two disks result in 33% higher performance, while three disks add an additional 23% (64% from one disk). Adding the fourth disk actually results in a drop from three disks, coming it at 5% lower than three and 56% higher than one.
Additionally, we also see that splitting disks across storage accounts makes no appreciable difference.  Note: Bar charts for this results section have been combined into the graphs below.

Disk Striping Methodology Tests:

  • 2 Disks, Striped Volume 32k Sector Size
  • 2 Disks, Storage Spaces Simple, 32k Sector Size
  • 3 Disks, Striped Volume, 32k Sector Size
  • 3 Disks, Storage Spaces Simple, 32k Sector Size
  • 4 Disks, Striped Volume, 32k Sector Size
  • 4 Disks, Storage Spaces Simple, 32k Sector Size

Table 4-Legacy Striping vs. Storage Spaces Test
Now we compare legacy striping to the newly introduced Storage Spaces. Two disk scaling is a definitive win for Storage Spaces, while beyond that legacy striping generally performs better (save max latency). In my opinion two disk Storage Spaces stripe is the sweet spot here (56% IOPS improvement!) when considering that the with more disks we add complexity that doesn't pan out on the performance side.


I hope you have found these results interesting; I certainly have. Even if you choose not to run these tests yourself I hope my results prove helpful when sizing your machines. Since the access pattern I used is relatively universal it should be applicable in most scenarios.
Software level disk striping works relatively well in Microsoft Azure to increase per-disk performance in lieu of a provider level solution similar to Amazon EBS provisioned IOPS. Splitting the workload across logical disks or VMs is preferred but not applicable to all workloads. When employing this solution make sure you select only locally redundant replication because Micrsoft warns that geo-redundant replication may cause data consistency issues on the replication target.
For additional information see the links near the top of this article. Thanks for reading!

Tuesday, May 20, 2014

Cloudy I/O Performance - Deciphering IOPS in IaaS (Part 1 of 2)

Note: This is part 1 of a 2 part post. Part 2 can be found here.


Disk performance scaling options in the public cloud seem limited (particularly in Azure as of this writing), but there are ways to increase your IOPS in IaaS solutions. To add to the performance problem, transactional costs of running application tests can be not only time consuming but expensive. To tune your storage performance reliably you will need a fast, consistent way to test different configurations. This article will cover that methodology and lead into a results/guidance article for Azure (but applicable to others) IaaS storage performance.

We'll be doing this testing on Windows, but you could also easily do this on Linux and the results that I'll be sharing are just as applicable there. To accomplish this testing we'll be using the following tools:

Let's begin!


We will proceed in the following order:
  1. Analyze Workload
  2. Create Test Scenarios
  3. Collect and Analyze Results (Mainly in Part 2)
  4. Findings (In Part 2)


If you plan on emulating my tests you'll need to have access to the following:
  • Microsoft Windows Azure account (note this methodology will worth with EC2 or any other platform including standard hardware/on-prem VMs)
  • IaaS VM Configured. A medium size is recommended for testing 4 disks or fewer to limit the available memory. More on that below.
  • Administrator access to your VM.
  • Your workload is in fact disk I/O bound. If you're not sure of that you may want to start with this article.
  • Awareness that you will incur additional storage transaction costs by running these tests.

Analysis/Create Workload

Note: If you're just trying to get a general sense for your VM I/O performance capability, you don't need to collect data for a custom access specification. IOMeter includes several tests you can use so skip to the "Install IOMeter..." section below.

The first thing we need to do is create our workload. By using IOMeter we can develop custom access patterns that model common workloads and have the tool and workloads installed and configured in minutes on any machine. There is nearly endless information on this topic, so I won't attempt to create a definitive source here. For details on how to configure and use IOMeter, see the following videos/articles:

 To create an accurate workload you will need a good understanding of the access pattern of your application. If you don't have that information you can use a tool like Perfmon to do analysis on a fully configured platform. The following counters will be of interest when creating your access specification:

  • Physical or Logical Disk: Average Disk Bytes per Read
  • Physical or Logical Disk: Average Disk Bytes per Write
  • Physical or Logical Disk: Disk Read Bytes/sec
  • Physical or Logical Disk: Disk Write Bytes/sec
  • Physical or Logical Disk: Disk Reads/sec
  • Physical or Logical Disk: Disk Writes/sec

For further information, see this excellent Technet Article.

By collecting this data during the access pattern you wish to emulate you can accurately estimate (with one caveat) the information needed to create the IOMeter access specification. That caveat is determining the sequential vs. random access pattern of the platform since Perfmon analysis will reveal the rest. To determine that, you'll need an understanding of how the platform stores and accesses/writes data. In my case I'm tuning my VM for Splunk, which uses a Map/Reduce functionality that has a highly sequential read/write pattern. If you are unsure of your access pattern then err on the side of configuring for mostly random access (90% or so) since it is generally more common and demanding of the underlying storage subsystem. 

Install IOMeter and Config Access Specification

The following actions can be done on your target testing platform or a different machine to stage settings. We'll be saving our settings for quick use later.

  1. Download and install IOMeter on your server. There are a series of ways to stage files on any VM, but if you're looking for a quick way in the Microsoft ecosystem check out my Onedrive/Azure post.

  2. Open IOMeter as administrator.

  3. Under "Topology" configure your workers. Each worker represents one thread generating I/O. By default it will create one per CPU thread available, but in most cases you will only want one worker per process you are emulating. In my case I'm assuming one large query at a time (and we'll scale from there), so I'll be testing with one worker. If you are unsure stick to one worker and you can move up from there when you become more familiar.

  4. Under "Disk Targets" select the disk you wish to test. This can change in later runs so if the disk you want to test isn't present here select a placeholder.
  5. Under "Disk Targets" configure your "Maximum Disk Size". This configures the size of your test file in sectors, which are considered to be 512 bytes each. To lessen the impact of OS caching you need to ensure this value exceeds the amount of RAM present on the machine to be tested. In my case I'll be testing on a 6GB RAM machine with a (approx) 7.5GB file, so I've configured it for 15000000 sectors. (15000000 sectors * 512 bytes per sector=7,680,000,000 bytes)  To do this quickly take your total desired size (in bytes!) and divide it by 512. (If you aren't certain you got it right, check the size iobw.tst file created at the root of your target drive after the first test is complete)

  6. Testing T: With a 4.5GB Test File

  7. Under "Disk Targets" configure your maximum outstanding I/O. This varies depending on access spec and OS, but I've had consistent (with real application access) results testing with 16 maximum outstanding I/O on windows. 
  8. Under "Test Setup" configure your "Ramp Up Time" and "Run Time". Ramp up need only be about 20 seconds for most scenarios and run time is best between 1 and 10 minutes. My results are based on (many per config) 5 minute tests. 
  9. Under "Access Specification" select your access spec. There is far too much to get into here; either select one or many existing access specifications that suit you needs ("4k 75% read" is a good start if you don't care) or create your own based on your findings from the Analysis/Create workload section. For the purposes of my test I made a "_Splunk" access spec with the following characteristics ascertained from my earlier performance testing:
    1. Transfer Request Size: 32kB (NOTE: My access spec may not reflect yours. Most won't be this large)
    2. Percent Read/Write Distribution: 53% Write/47% Read (NOTE: My access spec may not reflect yours. Most specs won't be this write heavy)
    3. Percent Random/Sequential Distribution: 75% Sequential/25% Random (NOTE: My access spec may not reflect yours. Most specs won't be this sequential)

  10. Add your access specification to the list of queued tests if you haven't done so already (removing all others).

  11. Click the disk icon to save the settings to an ICF file. This file will save all your settings including custom access specifications if applicable. Since this file is what you'll use to shortcut future testing, save it somewhere easy to transfer to other VMs such as OneDrive, Dropbox, SpiderOak, etc.

Run the Test

After setting up or loading your test settings, all you need do is click the green flag to start the test and then select where you would like to save the results. Make sure you don't overwrite any previous results and give the file a meaningful name so you remember what this test represents later, i.e. "results_3disk_1_StorAcct_Striped_32k_sectors_noCache_run1.csv" or similar. 

The test will run for the configured time and then you will be able to run additional tests or analyze results. Since the output is in CSV format, the natural place to look at this data is Excel. When IOMeter starts for the first time on a given disk it needs to create the test file. This will take quite awhile in both Amazon EC2 and Azure. (15 mins for my 7.5G for example) I believe this is due to the way space is allocated on the backend storage. Once this is created, however, you can run subsequent tests on the same volume without needing to wait for the test file to be created. Once the run is done I recommend running several more to ensure your tests aren't subject to wild performance swings. More on analysis in part 2 of this article.

How Much Will This Cost?

Since you're charged by transaction I'm sure you will be wondering how much this will cost. Let's break down your above baseline (system running) cost in Azure:

IOPS are currently capped at 500 for standard tier machines (300 for basic). Storage transactions are currently $0.01 per 100,000. (halved on 3/14/14) For every 5 minute test per disk you access you will then execute a maximum of 150,000 transactions. As a one time per configuration cost, you will need to build the test file which will be (test file size/volume sector size) transactions. For example, a 7.5 Gib test file will be approximately 1,875,000 transactions assuming a default 4kb sector size. (7,500,000,000/4000)

Test transactions + creation transactions = 2 million or so IOPS, or $0.20 @ .01 per 100,000. So... not much. The amount is generally trivial on Amazon EC2 as well. While this methodology will save you some in transaction costs, the main savings will be in time & labor. (which is usually our real cost anyhow!)

Further Optimization

Once you are comfortable with this process I would advise doing the following to further optimize this process. After doing so you may be able to automate the whole routine!

  • Create standard Perfmon counter sets for disk access and save/import them as a template
  • Script the Perfmon analysis with PowerShell
  • Create or download IOMeter templates for common access routines and include them with your set.
  • Script the installation and running of IOMeter, including multiple runs and uploading results to a common location. Easy to do with PowerShell and refer to the IOMeter manual for command line options (page 75 or so).
  • Package up all your assets with a custom installer and put it in an easy to get location. (mmmm... Chocolatey)
  • If you want angry followers and think digital bits are out there to be wasted, auto tweet your results! (maybe not this)

In Closing

I/O testing in the cloud is certainly feasible but requires a little extra discipline. With several access specifications in your toolkit you can conquer most performance problems quickly. What to do if your cloud platform doesn't provide your desired IOPS? Coming up in part 2!