addprocs() with SSHManager fails to load more than 1 process on remote node #18600
Closed
laurentheirendt opened this Issue on Sep 20 · 2 comments
Projects
None yet
Labels
parallel
Milestone
No milestone
Assignees
No one assigned
3 participants
@laurentheirendt
@ViralBShah
@tkelman
Notifications

You’re not receiving notifications from this thread.
@laurentheirendt
laurentheirendt commented on Sep 20

    Relevant versions: Julia 0.5.0-rc4 & v0.4.6 & 0.4.5 || Linux 14.04 LTS

Once the modifications are implemented as suggested in #18587 , the following command stalls:

cmd1 = addprocs([("<first.last>@server.com",2)], tunnel=true, sshflags=`-p9000`,topology=:master_slave, dir=`~/`, exename = "/usr/bin/julia/bin/julia", max_parallel=10);

whereas this works:

cmd2 = addprocs([("<first.last>@server.com",1)], tunnel=true, sshflags=`-p9000`,topology=:master_slave, dir=`~/`, exename = "/usr/bin/julia/bin/julia", max_parallel=10);

Moreover, cmd2 launches 16 julia instances with --worker in parallel. I guess that this is abnormal behaviour.

Testing shows that cmd1 stalls at line 1760 of Base/multi.jl

new_addresses = remotecall_fetch(launch_additional, frompid, cnt, cmd)

in the function launch_n_additional_processes(manager, frompid, fromconfig, cnt, launched_q).

@amitmurthy, do you know more about this?
@tkelman tkelman added the parallel label on Sep 20
@ViralBShah
The Julia Language member
ViralBShah commented on Sep 20

Cc @amitmurthy
@laurentheirendt
laurentheirendt commented on Sep 20

Version issue. All nodes and host upgraded to v0.5.0.
@laurentheirendt laurentheirendt closed this on Sep 20