[ofw] [ANNOUNCE] winOFED 2.3 RC3 available for download

[Chris Worley]

On Thu, Oct 7, 2010 at 11:27 AM, Tzachi Dar <tzachid at mellanox.co.il> wrote:
> Can you please explain what  "go nuts ...and ... locks up"?
> Is this blue screens hangs or anything else?
>
> Thanks
> Tzachi
>
>
>> In general, what versions of Windows are used in testing non-SRP parts
>> of the stack?  W2K8R2 "Standard" seemed to work well... but
>> "Enterprise" seems to go nuts w/ interrupts and NUMA distribution, and
>> occasionally locks up.

Sorry for the technical jargon ;)

I wasn't able to root cause anything, and am currently downgrading
back to "Standard", so everything is lost.

The first noticeable difference was benchmarking using IOMeter with 4
drives (all multipathed) formatted for NTFS each with one worker and a
QD of 256 and getting ~250K 4KB write IOPS.  After the switch to
Enterprise, the Windows console/GUI would become unresponsive (or
"intermittently responsive")... i.e. most windows would take a few
minutes to respond, including the ioMeter window... but monitoring the
target w/ iostat would show the expected performance.  This problem
would come and go.  In looking at taskmgr, it would show CPU 1 (the
second CPU) on the second NUMA node would be pegged at 100%, even if I
assured the IOMeter Dynamo's would not use that CPU.  All user space
tasks would show <11% CPU use, and I have no clue how to see what
kernel thread use is in Windows... but system idle was showing ~90%,
so I'm guessing there were interrupt floods hitting this one CPU.

After two days of this issue coming and going, the system started
locking up during the test, and became totally unresponsive, and
iostats on the target would show no activity.  The Windows console
would just show either the login screen (but mouse wouldn't move and
keyboard would not wake the screen) or just slow a screen-blanked
black screen.

That's as much detail as I can provide.

Sorry/thanks,

Chris