cache-quad. Here is my qsub: qsub -n 1 -t 80 --attrs knl_memory_mode=Cache -q knl_7250 donompi-m0_pn Steve Steven C. Pieper, Senior Physicist Emeritus: [email protected] Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100 http://www.phy.anl.gov/theory/staff/scp.html On 12/19/16 11:33, Allen, Benjamin S. wrote:

With which KNL mode(s) are you seeing this?

Ben

...

On Dec 19, 2016, at 11:22 AM, Steven C. Pieper <[email protected]> wrote:

The JLSE KNL nodes (I've tried knl09 and knl10) are about 40% slower for me after this change. This is true for both running load modules made before the change and running new modules. The slow down is fairly uniform in all the computationally intensive (Intel Fortran compiler) subroutines that I individually time. This does not appear to be my deallocate problem because some of the subroutines do no explicit allocates or dealloctes. It is also not a communication problem because I am running a 1-node non-mpi job. I am using OpenMP with 10 threads.

Have others compared speeds before and after the change?

Steve

Steven C. Pieper, Senior Physicist Emeritus: [email protected] Argonne National Laboratory, Physics Division, Bldg. 203, Argonne, IL 60439 Phone: 630-252-4232 Fax -6008 Secretary, Debra Beres, -4100 http://www.phy.anl.gov/theory/staff/scp.html

On 12/15/16 16:32, Allen, Benjamin S. wrote:

...

The OS image for the KNL machine has been updated to RHEL7.3 with kernel 3.10.0-514.2.2.el7.x86_64. RHEL7.3 is the first release where RedHat states it has KNL support. OmniPath OPA software has been updated to 10.3.0.0.81.

Hopefully these fixes will help with the various recent crashing conditions on these nodes.

Let me know what's broken.

Ben _______________________________________________ Intel-nda mailing list [email protected] https://lists.jlse.anl.gov/mailman/listinfo/intel-nda