[R] network connectivity is *not* symmetric -- library(parallel)
minshall at umich.edu
Thu Feb 14 15:08:49 CET 2013
hi. i've been very happy using multicore functionality, but felt the
need to get more horsepower, so have been trying to use parallel (after
realizing it is the "new snow") to access an amazon EC2 cluster, but
have had no success.
in the internet today, if machine M (the master) can open a connection
to machine S (the slave), it's somewhat of a miracle, something to be
celebrated. having accomplished that miracle, expecting that machine S
can *then* open a connection back to machine M is expected p^2 to be
close to p.
i get failures stating that the slave can't connect to the master, that
port 10187 "cannot be opened".
it would be nice if there were a mode whereby the *slave* would set up
to listen on a socket, transmit the socket number back to the master
(over the ssh connection the master has set up to the slave), and allow
the *master* (with proven ability to connect to the slave) to do the
would such a mode be possible? i suspect it would make a lot of people
happy. (network failures are pretty mysterious.)
and, while looking at the code, i note that the socket (port) number
being opened is, essentially, hard coded into parallels/snow. and, it
looks as if socketConnection requires a specification of the port
number. it might be nice to allow the system to choose the port number
[at the socket(2) call], and then report that back to the other side.
this would probably involve changes to socketConnection as well as maybe
the internet module, so is probably longer term. i think, though, that
doing this would benefit people who now are confused when they can't
start up two clusters on the same (currently, master) machine, etc.
cheers, Greg Minshall
More information about the R-help