Re: random i/o speedup weirdness

From: Jens Axboe <jens.axboe_at_oracle.com>
Date: Fri, 11 May 2007 13:56:29 +0200

On Fri, May 11 2007, Andras.Horvath_at_cern.ch wrote:
> Hi,
>
> I'm running fio 1.16 directly on a disk partition for an IOPS test.
> More specifically, I've created three partitions at the beginning, middle
> and end of the disk to verify some claims that the "beginning" of the
> disk is "faster", and to compare two disk types with regards to IOPS.
>
> So here I am running random read I/O agains a 4GB partition and,
> towards the end, the test speeds up considerably. This is true for all
> three partitions on two different direct-attached single disks.
>
> For example, running this:
> --- cut here ---
> [global]
> bs=4k
> direct=1
> rw=randread
> ioengine=libaio
> iodepth=2
> randrepeat=1
> write_bw_log
> write_lat_log
>
> [/dev/sdb1]
> --- cut here ---
> produces the attached throughput plot. Could someone please explain why
> it speeds up towards the end so significantly?
> (iostat actually confirms what fio reports)

Probably because fio gives up finding new free random blocks and then it
sort-of becomes a bit more sequential. Can you check if adding
'norandommap' to the options makes that effect go away?

-- 
Jens Axboe
Received on Fri May 11 2007 - 13:56:29 CEST

This archive was generated by hypermail 2.2.0 : Fri May 11 2007 - 14:00:01 CEST