[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[HTCondor-users] Collector load-balancing



Hello.

I'd like to ask for advice/feedback on scaling the collector of our glidein pool.

Some background: the number of worker nodes in the pool varies widely, but probably never exceeds 12k slots. Slot lifetime can be pretty short, so there is a lot of turnover. Many worker nodes are behind NATs and firewalls, so CCB is used. A pool password is used for authentication. Network latency is probably an issue. Lastly, our central manager is a VM with 8 virtual CPUs, and it uses shared_port.

Periodically, we've been observing spikes in numbers of log entries about timeouts, disconnects, ccb, shared_port failures and job restarts, so we've concluded that we need to run multiple collectors.

The plan is to follow https://htcondor-wiki.cs.wisc.edu/index.cgi/wiki?p=HowToConfigCollectors to create 5 additional collectors with custom socket names (since we use shared_port) and to configure glidein worker nodes to pick COLLECTOR_HOST and CCB_ADDRESS randomly among those 5 additional collectors.

Does anybody see potential issues with this? Or, maybe there is a better approach? Is there anything to be careful of?

Would it be advantageous if glideins used the same collector process for both COLLECTOR_HOST and CCB_ADDRESS? Or, maybe it would be advantageous to use some collectors exclusively for CCB_ADDRESS and other collectors exclusively for COLLECTOR_HOST?

I heard a few mentions of people running CCB on separate servers, but I am not sure why. Are there advantages to this if the central manager has idle cores and isn't running out of ports?


Thanks very much,

Vlad