Messages in this thread | | | Date | Tue, 10 Jan 2012 06:54:48 -0800 | From | Arjan van de Ven <> | Subject | Re: [PATCH] x86,sched: Fix sched_smt_power_savings totally broken |
| |
On 1/10/2012 6:41 AM, Peter Zijlstra wrote: > On Tue, 2012-01-10 at 06:32 -0800, Arjan van de Ven wrote: >> >> a very good default would be to keep all tasks on one package until half >> the cores in the package are busy, and then start spreading out. > > Does that still make sense when there's strong NUMA preference? By > forcing stuff on a single package you increase the number of remote > memory fetches (which generally generate more stalls), also the memory > controllers need to stay awake anyway.
the memory controllers need to stake regardless of what you do; it's more a memory bandwidth kind of thing.
if you have an enormous numa factor (>= 10 or so), then you really need a completely different policy I suspect. Thankfully those are rare.
| |