some fixes to benchmark framework including weak scaling of wave equation #577
Conversation
|
|
| "--different-devices", | ||
| is_flag=True, | ||
| default=False, | ||
| help="Ensures there are no multiple ranks on one SYCL device", |
There was a problem hiding this comment.
| help="Ensures there are no multiple ranks on one SYCL device", | |
| help="Ensures there are not multiple ranks on one SYCL device", |
include/dr/mhp/global.hpp
Outdated
| @@ -129,7 +129,8 @@ inline std::string hostname() { | |||
| inline sycl::queue &sycl_queue() { return __detail::gcontext()->sycl_queue_; } | |||
| inline auto dpl_policy() { return __detail::gcontext()->dpl_policy_; } | |||
|
|
|||
| inline sycl::queue select_queue(MPI_Comm comm = MPI_COMM_WORLD) { | |||
| inline sycl::queue select_queue(MPI_Comm comm, | |||
There was a problem hiding this comment.
| inline sycl::queue select_queue(MPI_Comm comm, | |
| inline sycl::queue select_queue( |
It is OK to delete the comm argument so the recommended usage can remain: select_queue(). We don't support or test different comms anyway.
| @@ -666,7 +673,8 @@ int main(int argc, char *argv[]) { | |||
|
|
|||
| static void WaveEquation_DR(benchmark::State &state) { | |||
|
|
|||
| int n = 4000; | |||
| int n = ::sqrtl(default_vector_size); | |||
There was a problem hiding this comment.
It seems strange that the running time is only 1 second because 4000 * 4000 = 16M and default vector size is 2B, but I see you did an actual benchmarking run.
There was a problem hiding this comment.
It seems strange that the running time is only 1 second because 4000 * 4000 = 16M and default vector size is 2B, but I see you did an actual benchmarking run.
I need to rerun. I've checked this again and I saw that I had manually specified a size when creating my plot. It was smaller than default. Hence 1sec :( I need to wait with checking this until they devcloud will resolve Invalid account or account/partition combination specified problem.
Ugly hack is to add n/=4 line to change default just in this test. Nice way is to change size of this test in suite in python.
There was a problem hiding this comment.
Previous plot was made with problem size 10M, now I've put temporarily an ugly hack and problem size is 500M, it runs in 10sec and gives the following plot
| @@ -447,6 +461,7 @@ def multi_node(base): | |||
| base.vec_size = [vec_size] | |||
| base.reps = reps | |||
| base.weak_scaling = weak_scaling | |||
| base.different_devices = different_devices | |||
There was a problem hiding this comment.
I am not happy with the code I wrote here. The redundancy in adding arguments is just 1 problem. Now that we understand better what the runner and plotter has to do, I expect something much simpler could work. Something to think about in the future....
|
The appropriate problem size depends on the benchmark. We need a size that will work for performance measurement and a size that is good for a quick test. Using --default-vector-size to control everything will not work. Something like:
|
I have an idea of writing suites in just flat way - no |
No description provided.