[Beowulf] [OOM killer/scheduler] disabling swap on cluster nodes?
prentice.bisbal at rutgers.edu
Tue Feb 10 11:11:39 PST 2015
On 02/10/2015 02:02 PM, Mark Hahn wrote:
>> ?Do you know if remote swap is working well now? I remember it
>> working not
>> so good because it's a difficult problem, but I haven't tested it
> I suspect remote swap works better now, in part because systems have
> been moving more towards specialized allocators. (motivated by wanting
> to scale-per-core, for instance.) it was always possible to make
> network-swap robust: all it takes is enough partitioning of storage
> pools, so that you don't run into an allocation-swap-allocation gotcha.
> is net-swap really ever a good idea? it always seems like asking for
> trouble, though in principle there's no reason why net IO should be
> "worse" than disk IO...
Sometimes it can be faster than disk I/O. Some SANs have massive amounts
of cache, so that data you've 'swapped to disk' might still be in the
SANs cache. I doubt this would be the case in a large cluster, but I've
seen and measured this effect first hand.
However, for HPC, I would think swapping to a network disk would
introduce some of the performance hits we're trying to avoid, like disk
latency, and add network latency on top of that.
More information about the Beowulf