Different GPU L1,L2 access count upon running with CPU

[Abhinav Sharma]

Hi,

I'm using Macsim to simulate CPU+GPU traces. I'm getting a higher number of GPU L1,L2 accesses when I run 1CPU + 1GPU, as compared to 1GPU only. I'm not able to reason why this should be so. In my understanding, CPU-GPU interference shouldn't affect GPU private cache accesses. Can anyone please help me understand this?

I've attached the stat files (all are dumped when the GPU finishes the first time), and also the params.out file. I'm running zeusmp (from SPECCPU2006) and hotspot (from Rodinia 3.1). L*_MISS_GPU+L*_HIT_GPU is different, while INST_COUNT_CORE_1 is the same in both. I observed similar behaviour on almost all other GPU benchmarks also.

 Any help will be much appreciated. Thanks!

Sincerely,

Abhinav Sharma

Indian Institute of Science

[Abhinav Sharma]

Hi,

As follow up to the above issue, I tried running 1c1g simulations with the two traces given with macsim (vectoradd and xalancbmk).

I noticed that the GPU cache access count in memory.stat.out file come out to be different upon changing the order of the CPU and GPU applications in trace_file_list. I'm not able to find out the reason for this difference. Any help will be highly appreciated!

To reproduce the error, use the attached trace_file_list (uses the two traces given in sst-unit-test/traces) and params.in (a simple modification of params/params_hetero_4c_4g). 

Thanks!

Sincerely,

Abhinav Sharma