SQLIO Benchmark Testing SSDs

Views 2779

SQLIO is a tool available for download from Microsoft that you can use to help determine the I/O capacity of your disk subsystem. Brent Ozar gives a great little video showing you how to use and configure your test batch scripts, as well as how to analyse your results.

There is a  section on SQLServerPedia that takes you through “How To Import your SQLIO Results into SQL Server” , the code is available for you to download which creates a couple of tables and a stored procedure which works through your SQLIO results text file to allow you to more easily analyse the results.

When using these scripts I did come across a couple of small issues around decimal places and length calculations which I have amended in the SQLIO_TestPass table creation and made the corresponding changes in the stored procedure to collate the data to report on.

clip_image001

 

Now that we have briefly gone over what is SQLIO, how you can download it, having created our import tables we are ready to start our testing of SSD’s. As Stated on SQLServerPedia, you need to ensure that your tests are actually getting to the disk and not just staying in cache. This means you are wanting to flush the cache, so your test files need to be larger than your cache. While recently on a client site working on a hardware refresh project which entailed moving onto newly purchased SSD’s. The hardware configuration we are testing here is as follows:

Samsung 200 GB (SS160520 – CLAR200) Configured in 6 * (4+1) raid groups
Presented Volumes are striped across the above configuration
EMC CLARiiON CX4-960
SP Read Cache 1000 MB
SP Write Cache 8054 MB
Qlogic QMH2462 4GB FC

SQLIO Test 1 Configuration

  • 10 GB Test file
  • 64 Kb Block size
  • 8 Kb Block size
  • 4 Threads with 1, 8 & 64 Outstanding IO requests
  • 8 Threads with 1, 8 & 64 Outstanding IO requests
  • 16 Threads with 1, 8 & 64 Outstanding IO requests
  • 32 Threads with 1, 8 & 64 Outstanding IO requests
  • Individual test duration – 180 seconds
  • Random Reads
  • Random Writes
  • Sequential Writes

8 Kb block size test result

Random Reads

clip_image002

 

clip_image003

 

clip_image004

 

Random Writes

clip_image005

 

clip_image006

clip_image007

 

Sequential Writes

clip_image008

 

clip_image009

clip_image010

 

64 Kb block size test result

Random Reads

clip_image011

 

clip_image012

clip_image013

 

Random Writes

clip_image014

 

clip_image015

clip_image016

 

Sequential Writes

clip_image017

 

clip_image018

clip_image019

 

After completing the tests and getting the results as seen in the graphs above, as would normally be expected the Random Read results show the SSD’s outperforms the non SSD’s for both sets of test (8Kb & 64 Kb block sizes). But the Write results were less than impressive. From these first round results the client understandably was concerned. The table below gives the highest values. From these values we try and find the sweet spot for what can be expected.

Block Size (Kb) IO Type IOPS Throughput MB/sec Latency (ms)
8 Random Read Upwards of 30,000 Upwards of 250 65
64 Random Read Upwards of 12,000 Upwards of 700 196
8 Random Write Upwards of 3,000 Upwards of 25 1046
64 Random Write Upwards of 1,500 Upwards of 70 1878
8 Sequential Write Upwards of 1,600 Upwards of 12 1270
64 Sequential Write Upwards of 280 Upwards of 20 7810

SQLIO Test 2 Configuration

With the results received in Test 1, the Random Write was not acceptable. This second round of tests was put together and undertaken to determine the configuration that would provide the appropriate performance.

  • 10 GB Test file
  • 50 GB Test file
  • 100 GB Test file
  • 64 Kb Block size
  • 32 Threads with 1, 4, 8, 16, 32 & 64 Outstanding IO requests
  • 64Threads with 1, 4, 8, 16, 32 & 64 Outstanding IO requests
  • Individual test duration – 180 seconds
  • Random Writes

32 Kb block size test result

Random Writes

clip_image020

clip_image021

clip_image022

clip_image023

64 Kb block size test result

Random Writes

clip_image024

clip_image025

clip_image026

clip_image027

With this second test being solely about random writes the results with the different Threads, Outstanding IO requests against the different file sizes, the results are far better than first time round. With analysing the results it looks like the system should be able to maintain up to 3000 IOPS with around 170 MB/sec throughput.

File Size (GB) Block Size (Kb) Threads IO Type IOPS Throughput MB/sec Latency (ms)
10 32 16 Random Write Upwards of 1,100 Upwards of 70 497
32 32 Random Write Upwards of 1,200 Upwards of 70 902
64 64 Random Write Upwards of 3,400 Upwards of 212 1686
50 32 16 Random Write Upwards of 3,100 Upwards of 195 168
32 32 Random Write Upwards of 3,000 Upwards of 187 341
64 64 Random Write Upwards of 3,000 Upwards of 180 1449
100 32 16 Random Write Upwards of 1,200 Upwards of 75 461
32 32 Random Write Upwards of 1,400 Upwards of 87 814
64 64 Random Write Upwards of 2,700 Upwards of 170 1524

Conclusion

With the testing that was undertaken to benchmark the new storage architecture, working closely with the storage administrator and in contact with the vendor, the client’s new system should be able to achieve far better performance than the previous kit. After the go live I will be running some benchmarking tests against the old system to give a comparison for the upgrade. If you are considering upgrading your storage, you need to know what you are needing to achieve from the performance so you can get all of the appropriate kit to meet those requirements. Do not be rash in your decisions, but work closely with your vendors to determine the appropriate kit for you.  The performance received from the SSD’s is very good but you do pay for that performance. As another option and I am not promoting a product or vendor, but have you heard of FusionIO? This could be a suitable option for you. Brent Ozar has undertaken some testing against this technology and his blog is a good read.

I hope that this has been interesting and an informative read around what you can achieve with your systems and a little bit of testing. All comments are welcome.

Leave a Reply

Your email address will not be published. Required fields are marked *

Warwick

I am a Microsoft Data Platform MVP as well as a Microsoft Certified Master working as the Principal Consultant here at SQL Masters Consulting. When I am not working with the SQL Server Stack I like to get away to the Snow and spend time Snowboarding.

Search