Skip to content
Prev 56975 / 63424 Next

MacOS parallel::makeCluster fails

Hi Dominik,

from the output, the master process could not "listen" on the port where 
it expects a connection from the worker. We need to find out why. I'd 
recommend first to create a minimal reproducible example (and one that 
does not use future, only parallel, and a minimal number of threads, 
ideally just 2). Then I'd recommend to check if the problem still exists 
with R-devel. Then I'd check if the problem happens in all invocations, 
even after reboots, on a clean system, without many running applications 
- if it does, this is good news. Then you could post such example and we 
could help more - if we can reproduce on our system indeed we could 
debug, if not there could at least be more directed advice on how to 
debug on your side. What I'd do myself if I could reproduce on my system 
would be instrument R around Sock_listen in internet module to see 
exactly what has failed with which error. Maybe dtruss would help too, 
but instrumenting may be easier. The earlier problem you mention has 
never been diagnosed (it was only intermittent on the reporter's 
machine, we could not reproduce on our systems, and despite a lot of 
effort on our side and on the reporter's, we could not reliably 
diagnose). In principle, it could be some race condition in R (one has 
been fixed since the previous report), but especially if it is 
deterministic it would more likely be some OS limit on your system. You 
could of course try playing with OS limits, on the number of open files, 
etc, with changing the port number (port= option), etc, but I would 
recommend the systematic approach of debugging the cause.

Best
Tomas
On 6/4/19 10:45 AM, Dominik Leutnant wrote: