PAPI error on AMD Opteron 6180 (romulus)

[fgsong]

Hi,

I'm trying to measure a native event on an AMD multicore machine (our 48-core romulus). The event name is "CPU_TO_DRAM_REQUESTS_TO_TARGET_NODE:*". There is no error if I start less than 5 threads, but then PAPI_read() starts to report errors if I have more threads (e.g., 6, 12 threads).
And here is the error information: Error in PAPI_read: PAPI_ESBST. BTW, I'm using PAPI 4.2.0.

Do you know what would be the reasons for the error?

Thanks,
Fengguang

[vweaver1]

Hi,

I'm trying to measure a native event on an AMD multicore machine (our 48-core romulus). The event name is "CPU_TO_DRAM_REQUESTS_TO_TARGET_NODE:*". There is no error if I start less than 5 threads, but then PAPI_read() starts to report errors if I have more threads (e.g., 6, 12 threads).
And here is the error information: Error in PAPI_read: PAPI_ESBST. BTW, I'm using PAPI 4.2.0.

Do you know what would be the reasons for the error?

This is strange. It's possible that that's a northbridge event; in that case it can only be used by one thread per core. I would think you'd be able to work on more than 5 threads though, I think you'd get at least 8. Are you pinning the threads to a CPU?

[fgsong]

Yes, I tried to pin 6 threads to 6 numa nodes, but there is still the same error. 5 threads are OK.

Thanks,
Fengguang

[fgsong]

My mask was not correct.
OK, now I find that I could run 8 threads if they are bound to 8 numa nodes (or 8 memory controllers).
The error happens if there are more than 8 threads.

Thanks,
Fengguang

[fgsong]

Does it mean it is not supporting one thread per core?

Thanks,
Fengguang