[ILUG] Disk throughput on current generation server hardware

[Conor Wynne]

Niall O Broin wrote:
> I have a few HP DL380 servers, with 2 x quad core 2.66 GHz Xeons, 12
> GB RAM, and 8 x 300 GB 10K RPM SAS disks in a RAID6 array on a Smart
> Array P410i controller.  The servers were originally deployed with
> SLES10SP2 where they proved very unreliable - sustained disk I/O would
> kill them.
>
> After spending a long time with HP on support tickets, where we were
> told to do such inane things as update the ILO firmware (because of
> course that's intimately involved in disk I/O) we eventually installed
> SLES10SP3 on one of them and SLES11 on another, and they seem to have
> ceased dying, which is of course A Good Thing. However, the disk
> throughput is now a bit disappointing. The simple test we were using
> to exercise the disks was to run in a loop the following:
>
> dd if=/dev/cciss/c0d0p4 of=/opt/dd_test count=20000 bs=1M
>
> I know we could do much 'better' tests with bonnie or some such but
> our initial objective was to provoke the failures which we were seeing
> in normal use, and which we suspected were disk I/O related, and the
> dd test did that just nicely. It gave us a means of relatively quickly
> seeing if HP's latest suggestion was of any use.
>
> The dd test gave the following approximate throughput figures:
>
>
> OS release    Kernel version        Throughput
>
> SLES10SP2    2.6.16.60-0.21-smp    100 MB/s
> SLES10SP3    2.6.16.60-0.54.5-smp     50 MB/s
> SLES11        2.6.27.19-5-default    30 MB/s
>
> SLES10SP2 is of course the outright winner here - unfortunately that
> speed came at the cost of stability.
>
> The reason for this post is to get people's views as to what kind of
> throughput I should reasonably be able to expect on that class of
> hardware.
>
>
> Niall
>
Are these the latest kernels and controller drivers?
"rpm -q --changelog kernel |grep -i cciss" comes to mind....

The OSes are 32-bit or 64-bit? If 32, and 12GB, have you tried the
largesmp/hugemem kernels?

Conor.