You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Right now we have a fixed maximum memory usage per worker based on the instance type - and because this isn't accurate we end up assuming that only 30% of it is available which leads to worse performance than needed. Instead, we should have each job that is created, compute the minimum free_memory = begin GC.gc(true); Sys.free_memory end across all workers using min_free_memory = MPI.Reduce(min, free_memory, 0, MPI.COMM_WORLD) and on the head node (where the MPI.Comm_rank(MPI.COMM_WORLD) == 0) we should send this to the client and then send this max_worker_memory to each call to evaluate which stores it in JobInfo.
The text was updated successfully, but these errors were encountered:
Right now we have a fixed maximum memory usage per worker based on the instance type - and because this isn't accurate we end up assuming that only 30% of it is available which leads to worse performance than needed. Instead, we should have each job that is created, compute the minimum
free_memory = begin GC.gc(true); Sys.free_memory end
across all workers usingmin_free_memory = MPI.Reduce(min, free_memory, 0, MPI.COMM_WORLD)
and on the head node (where theMPI.Comm_rank(MPI.COMM_WORLD) == 0
) we should send this to the client and then send thismax_worker_memory
to each call toevaluate
which stores it inJobInfo
.The text was updated successfully, but these errors were encountered: