Decrease in application performance overtime; occasional spikes of major slowdown

[brycelelbach]

HPX applications running with the MPI parcelport in distributed at scale experience a linear degradation in performance as application runtime increases. Also, periodically, there are spikes of massive slowdown.

This can be demonstrated by running the future_hang_on_get_629 regression test with the MPI parcelport like so:

mpirun -x LD_LIBRARY_PATH="$LD_LIBRARY_PATH" -machinefile machinefile -bynode -np 16 /home/wash/install/octopus/intel-13.0.1-release/bin/future_hang_on_get_629_test --verbose --depth=5 --test-runs=0

This regression test spawns a recursive tree (with a certain number of children per node, in this case, we use the default of 8), and calls the standard busy work null_function on each node of the tree. The depth of the tree is, in this case, 5 levels deep (analogous to 5 levels of refinement). We run this test for a certain number of iterations (e.g. timesteps)

The behavior demonstrated by this test case is very similar to the performance issues that the Octopus 3D torus simulation encounters. Here are some graphs demonstrating the problem. They plot the timestep speed (e.g. timesteps/second) for each timestep; e.g. instantaneous speed. Note that the first two graphs are logscaled.

629 regresison test, logscaled
Octopus, logscaled
629 regression test
Octopus

[sithhell]

Since you explicitly mention the MPI parcelport, does this happen with the TCP parcelport as well?
Could you please try to increase the number of max requests for the MPI parcelport (hpx.parcel.mpi.max_requests)? The default is 256 which might be too less for your application.

[brycelelbach]

The problem does not appear to show up with the TCP parcelport, but I'm not sure - the TCP parcelport is significantly slower which makes it harder to get enough data to draw any conclusions.

I'll try the max_requests thing.

[brycelelbach]

The max_requests suggestion fixes neither issue.

[hkaiser]

Some preliminary analysis showed that this is probably caused by a growing number of allocated stack segments. It is still unclear why/when this is happening.

[sithhell]

Is this still a problem?

[brycelelbach]

Yes

[sithhell]

Do we still have this problem?

[hkaiser]

I'm pretty sure this is caused by memory fragmentation increasing overtime. I don't have any other explanation. So I'm not sure we can do anything about this.

[hkaiser]

This has been fixed. Please reopen if appropriate (see #1753)