valgrind likwid_topology

[GoogleCodeExporter]

*What steps will reproduce the problem?
user@machine:~/valgrind ./likwid_topology

*What is the expected output? What do you see instead?

Not the real topology, more precisely cpuidcpuid_topology.numCoresPerSocket is 
saying only one core on the machine instead of 4. Moreover, it says that there 
is 6MB of L3 cache instead of 8MB.

See attachment with topology, topology_valgrind (when executed with valgrind) 
and the cpuinfo file

*What version of the product are you using?
likwid 3.0 and valgrind 3.8.1

*Please provide any additional information below.

the kernel cpuinfo is correct and in attachment

Original issue reported on code.google.com by teamexagraph@exascale-computing.eu on 11 Oct 2013 at 1:52

Attachments:

• topology
• topology_valgrind
• cpuinfo

[GoogleCodeExporter]

valgrind is emulating a single core processor as far as I know.

As likwid extracts all information directly from the cpuid instruction, it 
depends on the implementation of cpuid within valgrind. cpuid may be stubbed in 
a simplified fashion in valgrind.

Other tools which extract the information from a kernel fs tree are not 
affected because this does not if using valgrind.

Could you check how cpuid is treated within valgrind?

Original comment by jan.trei...@gmail.com on 14 Oct 2013 at 11:02

[GoogleCodeExporter]

Yes you are right on valgrind. It emulates a particular processor which doesn't 
correspond to the good one (see attachments). Very interesting by the way. 

Since it's valgrind's fault I think the issue can be considered as resolved.

Original comment by teamexagraph@exascale-computing.eu on 14 Oct 2013 at 12:43

Attachments:

• cpuid_result
• cpuid_valgrind_result

[GoogleCodeExporter]

Thank you for your time

Original comment by teamexagraph@exascale-computing.eu on 14 Oct 2013 at 12:44

[GoogleCodeExporter]

Original comment by jan.trei...@gmail.com on 14 Oct 2013 at 12:47

• Changed state: WontFix