[Ksummit-2013-discuss] NUMA locality for storage

Chris Mason chris.mason at fusionio.com
Mon Jul 29 20:22:12 UTC 2013


Quoting Matthew Wilcox (2013-07-27 23:00:21)
> Another potential topic (or subtopic if we do a session on low latency
>    devices).  With lower latency storage, the cross-socket latency and even
>    bandwidth starts to become an issue. We have NUMA APIs for memory, but
>    they focus on allocating memory local to the thread. If you're writing new
>    data, then sure, write it to a device close to the thread, but if you're
>    accessing already-written data, then it seems to me that an API that asks
>    for a cpumask close to a range of data might be a good idea.
> 
>    Or maybe not. Does this interact with anybody else's plans?

The goal should be memory and IO from the same node right?  So another
way to say it is give me memory and place my process on a CPU close to
device XYZ (this counts for networking too).

There isn't a ton of flexibility about where the device lives, unless
where on multipath.

NUMA aware swap device selection is also interesting.

-chris



More information about the Ksummit-2013-discuss mailing list