SFU Tests

Introduction

Tests

Hardware:

Worker nodes - b405, b408-414, all X5650 - 6 cores, 2 GB/core, total 96 cores

Similar to June2013 Tests

• The cluster is ready in exactly the same configuration as we had in June. The only difference is:
  • the WNs are running SL6.4
  • Torque server (bugaboo-dev.westgrid.ca) is running SL6.4 and Torque-4.2.6
  • The CE and CVMFS configurations are unchanged since June 2013.)

just a comment on "exactly the same configuration":
when the benchmarks were ran the last time there was a misconfiguration
in torque that had the effect that only half of the cores on the
computenodes were used: on the 12-core nodes only the cpus
0,2,4,6,8,10,12,14,16,18,20,22
were assigned to jobs. The problem is that all cpus with ids > 11 are
actually hyperthreads running on the real cpu id-12. Hence in the benchmark
in June we basically ran two jobs on a single core resulting in half of
the performance that we should have seen.
This is fixed now and all jobs should use cpus in the range 0 - 11.
The cpuset will still contain the hyperthread as well, but you need to
specifically change the way the job is running in order to use it.
E.g., if you submit a job requesting a single core (the default) the
cpuset will contain something like 4,16 (16 is the hyperthread belonging
to the real cpu 4), but unless you run a multithreaded job or a two
process MPI job that hyperthread will remain unused.

Test 20028852

• Test 20028852
• Dataset and Test Template: Description
• User analy job
• Direct access i/o used (site default)

Test 20028908

• Test 20028908
• Dataset and Test Template: Description
• PD3P production job
• copy-to-scratch used (overriding site default of direct i/o)

Results

-- AsokaDeSilva - 2013-11-20