Proper ClusterManager config for SharedArrays in SGE environment

[juliaparallel]

Proper ClusterManager config for SharedArrays in SGE environment

[pearcemc]

Hi. I ran into a similar issue and made some code in a separate package: https://github.com/pearcemc/ClusterUtils.jl . The function describepids and returns a dict with keys corresponding to the lowest numbered process on each node and values being lists of the processes on those nodes (default excludes master node). I also have some code somewhere for ensuring the resulting arrays are named everywhere they need to be - I can fish it out tomorrow if that would help?

[klkeys]

@pearcemc I couldn't get the ClusterUtils code to work, so I filed an issue on the repo.

But while I have your attention, is the premise behind describepids that I can use the process IDs that it returns for initializing the SharedArrays?

[pearcemc]

Thanks for filing the issue there. Yeah, that's the idea it provides the id of a representative process on each machine as a key to a list of the procs on the machine - which can be passed as the pids argument of a `SharedArray'. So you can create an appropriate array on each machine by iterating over the keys. The representative processes can be useful if you're doing blas operations as these use implicit parallelism, and when message passing (pass messages between representative procs only and shared mem takes care of the rest).

EDIT: My workflow for getting a SharedArray on each machine looks like.

@everywhere using ClusterUtils
T, V = (19,23)
sow(:T, T)
sow(:V, V)
Y_init = randn(T,V)
topo = describepids()
reps = collect(keys(topo))
sow(:topor, topo);
Yrefs = sow(reps, :Y, :( copy!(SharedArray(Float32, (T, V), pids=[topor[myid()]]), $Y_init) ));
for r in reps 
    sow(topo[r], :Y, :(fetch($Yrefs[$r])));   #this makes every proc on machine r request the SharedArray from its rep
end

Don't know how useful this will be to you due to versioning, but it was painfully acquired ;) If you find another solution I'd be interested to hear about it!

[juliohm]

Too old to reproduce. Please consider the latest release of the project.