Hi Peet âAs an aside, is there any documentation/literature for the maximum no. of nodes/slots a Windows and/or Linux CM can support?â Not sure what the limits are, but we have a windows only execute pool/s that all have linux (Ubuntu 20.04) CMs. These VMs have 2 cores and 4Gb RAM. All our submit nodes are Windows Server 2016 with 8-cores and 32Gb RAM. Total pool numbers = 1,900 machines with 10,000 slots/cores Our largest individual pool = 630 machines with 3,300 slots/cores All CMs also forward info to a condorview âsupercollectorâ but this is just to get total numbers and is only running a collector (no negotiator). Cheers Greg From: HTCondor-users <htcondor-users-bounces@xxxxxxxxxxx>
On Behalf Of John M Knoeller The 1024 comes from the compatibility layer that we use to make the Windows and Linux code bases the same, so we canât make it a knob unfortunately. We would have to change the Windows code base to use a more Windows
style mechanism for detecting hot sockets in order to go beyond this limit. -tj From: HTCondor-users <htcondor-users-bounces@xxxxxxxxxxx>
On Behalf Of Peet Whittaker 1014 comes from a hard coded limit of 1024 file descriptors for HTCondor on Windows (with a cushion), and there's no easy way to increase that, I fear. Ah, fair enough. I wonder whether this could be made into a config value in the future, as I think Windows can actually support a greater number of file descriptors (up to ~8k); see:
https://docs.microsoft.com/en-us/cpp/c-runtime-library/reference/setmaxstdio?view=msvc-160 But yes, we will look to migrate our CM to Linux (well, just the collector/negotiator parts anyway
😉). As an aside, is there any documentation/literature for the maximum no. of nodes/slots a Windows and/or Linux CM can support? I did find the following page, but it mainly talks about RAM requirements (there is nothing
related to file descriptors f.ex.) and is obviously quite old now (2007): https://research.cs.wisc.edu/htcondor/CondorWeek2007/large_condor_pools.html Kind regards, Peet Whittaker Discipline Lead for DevOps | Principal Software Developer From: Greg Thain <gthain@xxxxxxxxxxx>
On 5/31/22 16:23, Peet Whittaker wrote:
Note that the CM in our terminology is just the collector and negotiator process, not the schedd. I assume that if you split the collector & negotiator onto a linux machine, the other processes on your CM would stay on the Windows schedd/submit
machine? 1014 comes from a hard coded limit of 1024 file descriptors for HTCondor on Windows (with a cushion), and there's no easy way to increase that, I fear. -greg JBA Consulting, 1 Broughton Park, Old Lane North, Broughton, Skipton, North Yorkshire, BD23 3FD. Telephone: +441756699500
|