It'd be very helpful to have an idea of what the performance drop is - metadata
performance, and if so, what operation(s), and if dats performance, read or write and
other related details? Though, regardless, it's not very likely to be related. In
fact, you may find in trying to get hard details from the user the reported problem fades
From: HPDD-discuss [hpdd-discuss-bounces(a)lists.01.org] on behalf of Martin Hecht
Sent: Tuesday, August 11, 2015 4:52 AM
Subject: [HPDD-discuss] Lustre 1.8.9 wc1 for Kernel 2.6.32-573
We have built the 1.8.9 client for the latest rhel 6 kernel using the
patches mentioned before on this list.
Especially, we have also added the patch for LU-3596, but we have
received a complaint by one(!) of our users about a tremendous
performance drop when we applied the patch for LU-3596 and switched to
kernel 2.6.32-573.el6.x86_64 on the client side.
Since he is the only one complaining, it might be that he has introduced
simultaneously something else in his workflow, but has anyone heard
about performance issues related to the changed locking in the patch for
On 03/24/2015 09:40 PM, Chris Hunter wrote:
Yes, we still use Martin Hecht's backport of LU-1337 patch.
LU-5237 is used for builds with rhel 6.5 (it patches file
./kernel_patches/targets/targets/2.6-rhel6.target) and likely not
relevant for rhel 6.6.
ATM, we have o2iblnd issues for b1_8 clients with rhel kernels newer
than Oct 2014. IMO, its less work to move to a newer client version
than keep adding patches to b1_8 branch.
yale hpc group
On 03/24/2015 03:46 PM, Marcin Dulak wrote:
> On 03/24/2015 06:45 PM, Chris Hunter wrote:
>> We apply patch for LU2800 & LU5237 (b2d09eb6,diff?) to git b1_8
>> branch. A couple optional patches for LU3067 & LU3596 are available.
>> It has become difficult to build 1.8 clients with recent redhat
>> kernels (they tend backport fixes from the 3.10 kernels).