After I further disable CONFIG_DISABLE_CPU_SCHED_DOMAIN_BALANCE option, seems things get better now. And thread now could be distributed across the cluster.

One more question here, seems it is interesting to see that when A7 and A15 both are running at same frequency, their calculation capacity seems the same, which is proved by this fixed loaded test case.

Is there anyone has any idea why it could happen? Shouldn't A15's calculation power is stronger than A7? Then why the fixed load program would be finished with same time for both CPU?

Thanks, Lei

On Thu, Jun 27, 2013 at 5:38 PM, Lei Wen adrian.wenl@gmail.com wrote:

Sorry for inconvenience, as my bad mail editor... Change the mail subject to the corrected one

On Thu, Jun 27, 2013 at 5:34 PM, Lei Wen adrian.wenl@gmail.com wrote:

...

Hi Morten and list,

I am current investigating HMP related stuff over TC2 platform, and find a strange issue. Here I have two module, both are creating fixed loaded thread, but one module would bound those thread to the cpu, while another don't do the bound operation.

[I test with linaro 2013.05 release to get below result] With setting A7/A15 to the same frequency, 1G, The result is: Bounded one: Five thread finished with 137s/137s/137s/138s/138s

unbounded one(With HMP related configuration enabled, which is the default one): Five thread finished with 138s/275s/275s/275s/275s

unbounded one(With HMP related configuration disabled): Five thread finished with 228s/229s/229s/229s/231s

So it seems to me, that current configuration don't make TC2 to run in its full performance. It worries me for it may downgrade the benchmark for somehow.

I haven't see into details, just post the result here to get your guys' feedback. While I get more detailed analysis, I would also post it here.

Thanks, Lei