[ewg] Re: [Scst-devel] SRP performance in SCST

Bart Van Assche bart.vanassche at gmail.com
Mon Mar 10 07:53:46 PDT 2008


On Mon, Mar 10, 2008 at 3:42 PM, Doron Shoham <dorons at voltaire.com> wrote:
>
> Bart Van Assche wrote:
>  > Can you please try to use direct I/O instead of buffered I/O ?
>  >
>  > Bart.
>
>  I have tried to use dio but when the bpt was 128 (or higher) I got the following error:
>
>  echo 1 > /proc/scsi/sg/allow_dio
>
>  sgp_dd if=/dev/sg1 of=/dev/null bs=512 bpt=128 thr=8 time=1 count=5M dio=1
>  time to transfer data was 2.236290 secs, 1200.36 MB/sec
>  5242880+0 records in
>  5242880+0 records out
>  >> Direct IO requested but incomplete 40960 times

I'm not familiar with the /proc/scsi/sg/allow_dio entry. I was
referring to options like "iflag=direct" or "oflag=direct" for dd --
all performance tests I have performed in the past were with dd and
xdd. Without direct I/O, read tests use Linux' buffering. If the
amount of data being read is less than the available RAM, buffered I/O
tests only measure how fast Linux can copy data in RAM.

Bart.



More information about the ewg mailing list