Ridiculous random write I/O with SAS SSD

[Robotbeat]

This is using the Atlanta build, I think. Build 3278 (pre-atlanta) shows similar performance, but may be about 10% slower, but I'm not sure.

Here are the results of sqlio in a system with only 1GB of RAM. This is for sustained, random 4k writes using an enterprise-grade SSD (that happens to use SAS and have over 120 GB of space and is 2.5"). This is for Block I/O. For File I/O and for Fibre Channel, there is slightly better performance for small test file sizes and the write iops drop to around 2000 for test files that do not fit in the cache. I think this has to do with the overhead of using file i/o (and I think the fibre channel target uses a type of file i/o).

[Robotbeat]

Here, you can see the results of the overhead of the linux page cache (i.e. file i/o):
(BTW, this test uses a 1Gb fibre channel link on the initiator side)
Random 4k Writes:


Random 4K Reads:


Still, even with the File system overhead, it's very fast.

So, what about Block I/O for fibre channel?