Qn on multi GPU - simpleMultiGPU Should main calling thread use/manage one device?

[winguy]

Hi,

From simpleMultiGPU.h :

• There’s one CUDA context per thread. To use multiple CUDA contexts you

• have to create multiple threads. One for each GPU. For optimal performance,

• the number of CPU cores should be equal to the number of GPUs in the system.

Assuming a system has a N-core CPU, and N GPUs, simpleMultiGPU will spawn N threads, 1 for each GPU. From my (little) understanding, there are now N + 1 host threads (main calling thread + N spawned threads) running on a N-core CPU. My question is, which is more optimal:

( A ) N GPUs, N + 1 CPU threads (inclusive of main calling thread)

or

( B ) N GPUs, N CPU threads (inclusive of main calling thread)

( B ) means managing one GPU device (e.g. device 0) on the main calling thread and creating new threads for subsequent GPU devices.

I looked at MonteCarloMultiGPU and it uses ( A ) N + 1 CPU threads (inclusive of main calling thread).

Thanks! :)

[eyalhir74]

Hi,

From simpleMultiGPU.h :

Assuming a system has a N-core CPU, and N GPUs, simpleMultiGPU will spawn N threads, 1 for each GPU. From my (little) understanding, there are now N + 1 host threads (main calling thread + N spawned threads) running on a N-core CPU. My question is, which is more optimal:

( A ) N GPUs, N + 1 CPU threads (inclusive of main calling thread)

or

( B ) N GPUs, N CPU threads (inclusive of main calling thread)

( B ) means managing one GPU device (e.g. device 0) on the main calling thread and creating new threads for subsequent GPU devices.

I looked at MonteCarloMultiGPU and it uses ( A ) N + 1 CPU threads (inclusive of main calling thread).

Thanks! :)

Hi,

You shouldnt manage the GPU on the main thread. Opening an additional thread (beyond the N cores won’t do any harm).

Just look at how many threads explorer, outlook, etc uses… :)

You should have one-one relationship between CPU threads and GPU devices.

eyal