The parameter is called "max_switch_wait" and has a default value of 300 seconds. See:
http://slurm.schedmd.com/slurm.conf.html Regards, Uwe Am 13.12.2014 um 16:16 schrieb Antony Cleave: > Yeah use the hostnames which slurm knows and it will work perfectly. > > You might also want to increase the max switch wait parameter in slurm > which offhand I cannot remember how to do but I do remember that the > default maximum is about 10 minutes which I consider to be a bit on the > short side if you have a nontrivial blocking factor and a long > interconnect bound application. It is documened however. > > Good luck! > > Antony > > On 13 Dec 2014 15:06, "Williams, Kevin E. (Federal C&I)" > <[email protected] <mailto:[email protected]>> wrote: > > Many thanks, Antony.____ > > __ __ > > As my config uses different names, from the slurm node names, for > the IB interconnect, it should be OK to use the ib topology but > identified by their slurm node names?____ > > __ __ > > If so, then we are set for testing.____ > > __ __ > > kevin____ > > __ __ > > *From:*Antony Cleave [mailto:[email protected] > <mailto:[email protected]>] > *Sent:* Saturday, December 13, 2014 9:09 AM > *To:* slurm-dev > *Subject:* [slurm-dev] Re: Quick question on use of topology.conf____ > > __ __ > > I think the quick answer is no.____ > > I will use an example of a system I configured recently to explain. > This system uses a 1Gb network for all slurm traffic and system > services and an infiniband network for interprocess communication. > As such I configured the topology file to be based on the infiniband > switch topology rather than the ethernet topology.____ > > Antony____ > > On 13 Dec 2014 12:53, "Williams, Kevin E. (Federal C&I)" > <[email protected] <mailto:[email protected]>> wrote:____ > > Researching the use of topology.conf and wondering if the switch > topology used must correspond to the interconnect that is used > for slurm communications. That is, if nodes in the slurm > configuration have multiple interconnects available, and the one > used to define the nodes in slurm partitions is a 1GB network, > does the switch topology have to reflect that particular network > topology?____ > > Image removed by sender.____ >
