Re: 答复：答复：md raid5 performace 6x SSD RAID5

[Stan Hoeppner <stan@hardwarefreak.com>]

On 11/28/2013 8:38 PM, Stan Hoeppner wrote:
> On 11/28/2013 4:02 AM, lilofile wrote:
>> thank you  for your advise. now I have test multi-thread patch, the single raid5 performance improve 30%.
>>
>> but I have another problem,when write on single raid,write performance is  approx 1.1GB/s 
> ...
>> [1]-  Done                    dd if=/dev/zero of=/dev/md126 count=100000 bs=1M
>> [2]+  Done                    dd if=/dev/zero of=/dev/md127 count=100000 bs=1M
> 
> No.  This is not a parallel IO test.
> 
> ...
>> To address #3 use FIO or a similar testing tool that can issue IOs in
>> parallel.  With SSD based storage you will never reach maximum
>> throughput with a serial data stream.
> 
> This is a parallel IO test, one command line:
> 
> ~# fio --directory=/dev/md126 --zero_buffers --numjobs=16
> --group_reporting --blocksize=64k --ioengine=libaio --iodepth=16
> --direct=1 --size=64g --name=read --rw=read --stonewall --name=write
> --rw=write --stonewall

Correction.  The --size value is per job, not per fio run.  We use 16
jobs in parallel to maximize the hardware throughput.  So use --size=4g
for 64GB total written in the test.  If you use --size=64g as I stated
above you'll write 1TB total in the test, and it will take forever to
finish.  With --size=4g the read test should take ~30 seconds and the
write test ~40s, not including the fio initialization time.

> Normally this targets a filesystem, not a raw block device.  This
> command line should work for a raw md device.


-- 
Stan