Fusion I/O Benchmark

[powellbound]

Anyone done any benchmarking with the Fusion I/O SSD boards??

Thanks!!

[David (PassMark)]

If someone would like to send us one, we'd love to have a play with one.

There are probably some example benchmarks for it already in our database, but I am fairly sure that the PCI card fails to identify itself as a normal drive, and so the name and description for the drive aren't being picked up automatically by PerformanceTest, making it hard to search for in the database of results. It is further not helped by several other products being on the market, also called "Fusion", and PerformanceTest testing Drive C by default.

But I did find one, Baseline number 185074. Read speed was 1056MB/sec. Write 914MB/Sec. Random Seek 403MB/sec. Which are pretty amazing numbers.

[Alonso]

Hi David,

Could you find out which settings were used in the PassMark in order to test the Fusion i/o drive and also what kind of drive was it?

This because I have some 320GB MLC fusion drives, installed on some servers with Windows 2008 R1 SP2 but Pass Mark shows its performance close to a normal C: drive (SCSI 320GB 15k)

For example on a normal SCSI drive with the follwing settings I get 170 MB/sec, while on the Fusion drive with the same settings I get 102 MB/secs.

100/0 R/W
50/50 Sequential/Random
File Size 500MB
Block Size 16384
Access Method – Std Win32 API (uncached)
Requested Mode – Synchronous


Any idea what am I doing wrong? Thanks so much for looking into this!!

[David (PassMark)]

I would suggest

1) Making the File size much larger. If the drive is reading at 1GB/sec, then a 512MB test file is tiny. (the file would be full in 0.5 sec, which isn't a very long test if you are looking a sequential performance). In some drives (solid state included) we see a latency effect. So they run a bit slowly for between 0.5 to 1 seconds, then start to pick up speed.

2) Try using a larger Block size. This tends to give better number for all drives. But small block are especially bad for solid state drives. The manufacturers of the SSDs normally only quote the best performance figures they can find. They don't tell you about the mediocre performance you might get with smaller I/O sizes.

3) Try switching the Asynchronous request mode. Which a queue length of 10 to 20. This reduces latency of sending commands to the drive and also allows the drive to reorder requests to be more efficient.

4) Try using PerformanceTest V8 as you cal also adjust the randomness of the data which impacts on the compressibility, which can impact on the speed of some SSDs which do file compression in the background.

5) Of course switching on caching will give better numbers as well. You just need to be careful not to end up measuring the speed of the RAM cache (unless that is your goal). If you use a very large File Size in the test (larger than the RAM in the machine) then this can force the file out of RAM and on to the disk.

[Alonso]

David,

Thanks for replying wich such detail so quickly. I will follow you good suggestions and post the results here soon.