GPU+nCPU

[go2master]

GPU+nCPU means run the data-partition benchmark with n CPU threads , and all with default partition in your paper?

[lchang20]

The number of CPU threads can be specified using -t. Without specifying, the default value is set as 4. In the paper, we showed the results of 1,2,4 CPU threads (with and without GPU) for an AMD Kaveri APU.

[go2master]

Is there no variations of partition ratio between GPU and CPU in the paper?

[lchang20]

The optimal ratios vary among benchmarks and datasets. The ratio can be specified statically using -a with a value between 0 and 1. Without specifying, the default value is set as 0.1. The ratio can be also dynamically determined if -a is specified out of [0,1]. In the paper, we showed the best static partition ratios for an AMD Kaveri APU.

[go2master]

Thank you.

[go2master]

And, when you send the patch for gem5-gpu simulator?

[el1goluj]

CUDA-D-Sim and CUDA-U-Sim are Chai's versions for gem5-gpu.
You can also find them in gem5-gpu benchmark repo (https://gem5-gpu.cs.wisc.edu/repo/benchmarks/).

[go2master]

I mean the patch to enable gem5-gpu execute Chai correctly. Victor said he forgot that.

[el1goluj]

gem5-gpu guys already received the patch. If it isn't pushed yet, they might be waiting for others to push several patches together.
We can email it to you personally. If you want so, please email me at [email protected].

Juan

[go2master]

Thanks a lot.