r - How to set up cluster slave nodes (on Windows) -
i need run thousands* of models on 15 machines (each of 4 cores), windows. started learn parallel
, snow
, snowfall
packages , read bunch of intro's, focus on setup of master. there little information on how set worker (slave) nodes on windows. information contradictory: some sock cluster practically easiest way go, others claim sock cluster setup complicated on windows (sshd setup) , best way go mpi.
so, easiest way install slave nodes on windows? mpi, pvm, sock or nws? my, possibly naive ideas (listed priority):
- to use 4 cores on slave nodes (required).
- ideally, need r packages , slave r script or r function listen on port , wait tasks master.
- ideally, nodes can added/removed dynamically cluster.
- ideally, slaves connect master - wouldn't have list slaves ip's in configuration of master.
only 1 100% required, 2-4 "would good". naive request?
i sorry have not been able figure out available docs , tutorials. grateful if point me out right source.
* note each of thousands of models take @ least 7 minutes, there won't big communication overhead.
it's shame how these apis (like parallel/snow/snowfall) complex work with, lots of docs not need... have found api simple , goes straight ideas sketched!! redis , doredis
r package (as recommended here). very simple tutorial present! modified bit , got this:
the workers need r, doredis package , script:
require(doredis) redisworker('jobs', '10.0.0.7') # ip of server
the master needs redis server running (installed the experimental windows binaries windows), , r code:
require(doredis) registerdoredis('jobs') foreach(j=1:10,.combine=sum,.multicombine=true) %dopar% ... # whatever need run removequeue('jobs')
adding/removing workers dynamic, no need specify ips @ master, automatic "load balanancing", simple , no need tons of docs! solution fulfills requirements , more - stated in ?registerdoredis
:
the doredis parallel end tolerates faults among worker processes , automatically resubmits failed tasks.
i don't know how complex using parallel/snow/snowfall socks/mpi/pvm/nws, if possible @ all, guess complex...
the disadvantages of using redis found:
- it database server. wonder if api exist somewhere without need install database server don't need @ all. guess must exist!
- there bug in current doredis package (
"object '.doredisglobals' not found"
) no solution yet , i not able install old working doredis 1.0.5 package r 3.0.1.
Comments
Post a Comment