On Sat, Nov 03, 2012 at 01:57:10PM -0400, erik quanstrom wrote: > > however, i think that queuing theory in general says that one queue with > global > sorting beats n smaller queues with local sorting. i think this is sometimes > called the > checkout-line problem. >
This would be true, I think, if the "big queue" had sufficient visibility into the guests. They almost never do. > other factors, like global knowledge of memory use stats and page duplication > should put the vm in an even better position than general queueing theory > would suggest to make decisions on what pages to move to disk wrt. global > (that is total machine) throughput. Page dedup, at least on linux, is not particularly useful for virtualization yet. It can be made to work but I've never seen people benefit from it unless the guest operating systems are also linux systems. > do you have a reference that demonstrates or derives that a similarly-loaded > machine can perform better with all the guests swapping indepdently and > the vm not swapping, rather than preventing the guests from swapping and > letting the vm swap? I do not have any formal data on the subject; only the things I've seen "in the field" as it were. 9cloud does not oversubscribe or page out VMs unless the system is in danger of crashing. vm.overcommit_memory is set to 2 and swap is there in case a system process spirals out of control. The systems are reasonably performant. I manage the hypervisor in this style based on years of seeing other options burn me.
