I have been doing a bit of testing recently with the LSI SAS 2008 based controllers such as the 9211-8i in RAID 0 with various SSDs. Recently I took eight 64GB ADATA S599 Sandforce based SSDs and placed them in RAID 0 on the LSI 9211-8i. I was expecting maximum transfer rates in excess of 2GB/s in ATTO. What I saw instead was a performance anomaly.
For the test configuration I used the same setup as I have been using for my SSD reviews:
- CPU: Intel Core i7 920
- Motherboard: Gigabyte X58-Extreme
- Memory: 12GB of Corsair Dominator GT 1600 C7 DDR3
- Case: CoolerMaster Cosmos S
- Drives (OS): 2x OCZ Vertex 120GB in RAID 0
- Controller: LSI 9211-8i SAS 2008 based controllers
- SSDs: OCZ Agility 2 120GB and ADATA S599 64GB
- Controller: Intel ICH10R with Intel Rapid Storage Technology (RST) 188.8.131.524
- NIC (additional): Intel Pro/1000 PT Quad
- Host OS: Microsoft Windows 7 Ultimate 64-bit
LSI driver version 2.00.35 was used in Windows 7 with the 07.00.00.00 firmware being used including the 07.11.00.00 mptsas2.rom included in the firmware download.
Using ATTO as the best-case scenario test for Sandforce based SSDs due to its use of compressible data I saw something completely unexpected. Eight drives were performing slower than two and much slower than four drives in RAID 0.
Clearly something is awry in the 128k and up test sizes.
After giving it some thought, I decided to change my strategy and try two RAID 0 arrays of four drives with each RAID 0 set being tied to a SFF-8087 port on the LSI 9211-8i. After that, I took the two RAID 0 sets and used a Windows Dynamic Disk stripe (another form of RAID 0) to combine the two sets. The results speak for themselves, over 2.1GB/s in raw bandwidth with just under 440GB usable capacity (this would be advertised as a 480GB solution depending on the manufacturer.)
This is in-line with what I expected to see from eight Sandforce SFF-1222 based SSDs in RAID 0.
Just to sanity check I tried this with a second LSI 9211-8i controller and a Supermicro X8SI6-F motherboard also with the same SAS 2008 controller and saw the same results so this is not a one-off controller specific issue. It should also be noted that the same type performance issues occurred in both AS SSD and CrystalDiskMark benchmarks so this was not ATTO specific, but ATTO shows a good illustration.
It is fairly well known that RAID 0 has very little CPU overhead, and it appears as though spanning a RAID 0 array across the two SFF-8087 controllers on SAS 2008 based controllers is a bad idea despite the 533MHz RAID controller. What can also be seen clearly from the results is that there is the raw performance potential for over 2GB/s with eight drives but there is something, probably in firmware, limiting the performance of the eight drive configurations. The main negative here is that switching to Dynamic Disks to get that performance means having a non-bootable array. Hopefully LSI will fix this in an upcoming firmware patch.