Backing up what everyone else says.  DO you know how the 24 core jobs are being 
distributed between sockets and cores.

Also as a very naive diagnostic, just run 'top' or even better 'htop' while the 
jobs are running.
I KNOW this isnt a great diagnostic for parallel programs, but sometimes you 
can notice patterns.
Select the option in top to show you the cores which are running.

Also keep an eye on the memory assignments at the top - you are not going into 
swap?
Are any processes going into D sttes for extended amounts of time?


________________________________________
From: Beowulf [[email protected]] on behalf of Peter St. John 
[[email protected]]
Sent: 29 February 2016 00:50
To: Jon Tegner
Cc: Beowulf List
Subject: Re: [Beowulf] Scaling issues on Xeon E5-2680

I come from the math side, not the electronics side, so this may be an 
ill-posed question, but could you try running the job with 12 cores on just one 
node, then with 6 cores on each of two nodes? I'm thinking, the 24 core version 
may get assigned to more nodes than your 12 core, and it's communication 
between nodes that slows it down, e.g. for a job that doesn't require so much 
memory as inter-process communication.
Peter

On Sun, Feb 28, 2016 at 11:01 AM, Jon Tegner 
<[email protected]<mailto:[email protected]>> wrote:
Thanks!

we are using openmpi-1.10.2, and I believe bind-to core and socket are on by 
default here. Openmpi is built using

./configure \
        --build=x86_64-redhat-linux-gnu \
        --host=x86_64-redhat-linux-gnu \
        --disable-dependency-tracking \
        --prefix=/remote/soft/OpenMPI/1.10.2 \
        --disable-mpi-profile \
        --enable-shared \
        --with-tm \
        --with-sge \
        --with-verbs

gcc is used throughout.

Code should be using RDMA. I have verified that infiniband performance (using 
openmpi-1.10.2) is what is to be expected (through mpitests-osu_bw and 
mpitest_osu_latency).

Hugepagesize haven't been touched ("grep Hugepagesize /proc/meminfo" gives 2048 
kB). Is this something worth changing?

numastat gives

                           node0           node1
numa_hit                 6288694         1193663
numa_miss                      0               0
numa_foreign                   0               0
interleave_hit             66539           66287
local_node               6288301         1126078
other_node                   393           67585

on one of the nodes - but these numbers seems to be fairly representative for 
the nodes. However, haven't used numastat before, and maybe it should be used 
simultaneously as the job is running (which it wasn't now)?

Thanks!

/jon


On 02/28/2016 04:30 PM, 
[email protected]<mailto:[email protected]> wrote:
Do you have CPU affinity enabled? Is this omp + MPI? Which compilers and flags. 
Give more details about the software stack

   Original Message
From: Jon Tegner
Sent: Sunday, February 28, 2016 22:28
To: [email protected]<mailto:[email protected]>
Subject: [Beowulf] Scaling issues on Xeon E5-2680

Hi,

have issues with performance on E5-2680. Each of the nodes have 2 of
these 12 core CPUs on SuperMicro SuperServer 1028R-WMR (i.e., 24 cores
on each node).

For one of our applications (CFD/OpenFOAM) we have noticed that the
calculation runs faster using 12 cores on 4 nodes compared to when using
24 cores on 4 nodes.

In our environment we also have older AMD hardware (nodes with 4 CPUs
with 12 cores each), and here we don't see these strange scaling issues.

System is CentOS-7, and communication is over FDR Infiniband. BIOS is
recently updated, and hyperthreading is disabled.

Feel a bit lost here, and any hints on how to proceed with this are
greatly appreciated!

Thanks,

/jon
_______________________________________________
Beowulf mailing list, [email protected]<mailto:[email protected]> sponsored 
by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit 
http://www.beowulf.org/mailman/listinfo/beowulf

_______________________________________________
Beowulf mailing list, [email protected]<mailto:[email protected]> sponsored 
by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit 
http://www.beowulf.org/mailman/listinfo/beowulf

#####################################################################################
Scanned by MailMarshal - M86 Security's comprehensive email content security 
solution.
#####################################################################################
Any views or opinions presented in this email are solely those of the author 
and do not necessarily represent those of the company. Employees of XMA Ltd are 
expressly required not to make defamatory statements and not to infringe or 
authorise any infringement of copyright or any other legal right by email 
communications. Any such communication is contrary to company policy and 
outside the scope of the employment of the individual concerned. The company 
will not accept any liability in respect of such communication, and the 
employee responsible will be personally liable for any damages or other 
liability arising. XMA Limited is registered in England and Wales (registered 
no. 2051703). Registered Office: Wilford Industrial Estate, Ruddington Lane, 
Wilford, Nottingham, NG11 7EP
_______________________________________________
Beowulf mailing list, [email protected] sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit 
http://www.beowulf.org/mailman/listinfo/beowulf

Reply via email to