[Beowulf] SSD caching for parallel filesystems

[Ellis H. Wilson III]

On 02/08/2013 11:50 AM, Jonathan Aquilina wrote:
> What would be the best way to bench mark these devices be it the sata
> ssd's or pcie. I have a sata based ssd in my netbook and its quite zippy.

Joe should have some nice suggestions here, but ultimately the best way 
to benchmark it for OPs case is going to be with real applications. 
Traditional I/O testers include:

- dd
- bonnie++
- iozone
- hdparm

But the issue with all of these is that, without working with large 
amounts of data and over long periods of time, the real "differences" in 
the SSD flash-translation layers doesn't come out.  If you are set on 
synthetically testing it, my suggestion is to first fill the SSD half or 
more than half with a mixture of random and sequential data.  Then, test 
sequential and random data, both block-aligned and otherwise, both at 
the same time and by themselves.  The best SSD solutions out there will 
continue to perform even as they approach filled and even for heavy 
random, unaligned, multi-stream accesses.  That's the real magic of the 
FTL at work, and is why these controller companies sell for many 
millions of dollars.

Best,

ellis