May 23, 2014
In my blog post that supported the announcement of the VSP G1000 with SVOS, Storage Virtualization Operating System, I described VSP G1000 as a unified storage platform that can provide approximately 4 million random reads for 8KB blocks or more than 1.2 million NFS operations for file.
While the 1.2 million NFS operations is noted in our online spec sheets, documented in a white paper and on the SPECsfs website, the approximately 4 million 8KB IOPs is not documented in a similar manner. The reason for this was that the review of the test results, inside and outside of HDS were not complete by the cut off date for the announcement. I was not aware of this when I created my post. Since the IOPs that I reported was 4 times the best performance of any other flash vendor, there have been questions of how I arrived at this performance number.
I am told that the report will be ready soon, so I would like to give you an update and prod our internal review process to get this officially published in our specifications.
The test was run in our labs with 100% random 8K reads generated by a multi-host vdbench. The configuration was as follows:
16 x Hitachi Compute Blade E55A2 running Red Hat Linux 6.4
32 x Emulex LPe16002B-M6 PCIe 2-port 16 Gb FC adapters (two HBAs/four paths per blade – directly connected to VSP G1000 at 8Gb)
VSP G1000 Dual Chassis “Power Configuration”
8 VSD Virtual Storage Director pairs (128 cores)
4 FED Front End Director pairs (64 ports)
4 CPC Cache Path Controller pairs (1 TB cache)
4 BED Back End Director pairs with 192 x 1.6 TB FMDs
The test team was able to get 3.9 million cache-miss IOPs with an average response time of only 1.6 ms with a standard thread count of 32 threads per FMD. With a thread count of 32 they needed all 192 FMDs to reach 3.9 M IOPS. With higher thread counts they were able to reach 4 million IOPs with as few as 64 FMDs.
How does this compare with other vendor’s specification? Here is a table that I put together based on vendor specs that are available on their websites:
The maximum IOPs differ by how they report it. In IBM’s case they show in their spec sheet that the maximum 1,000,000 IOPs comes from 100% cache hits while reads from cache misses (random reads) drops to 300,000 IOPS. EMC, IBM, and HDS report random reads which go to the backend flash drives. Other vendors do not specify how they measure their maximum IOPs. I have not crawled through all the different configurations, but I suspect that our G1000 and FMD architectures require fewer compute blades, FC ports, and flash drives than any of the other vendors while delivering almost four times the IOPs.
I am not ready to discuss how we achieved this performance breakthrough at this time. There are some hints in the configuration, where we describe 8 VSD pairs with 128 cores, and thread counts. Stay tuned for the white paper. Until then, please use this post for support of my performance claim for the VSP G1000.