[LSF/MM TOPIC][ATTEND]cold page tracking / working set estimation

[LSF/MM TOPIC][ATTEND]cold page tracking / working set estimation

[Michel Lespinasse]

Google uses an automated system to assign compute jobs to individual
machines within a cluster. In order to improve memory utilization in
the cluster, this system collects memory utilization statistics for
each cgroup on each machine. The following properties are desired for
the working set estimation mechanism:

- Low impact on the normal MM algorithms - we don't want to stress the
VM just by enabling working set estimation;

- Collected statistics should be comparable across multiple machines -
we don't just want to know which cgroup to reclaim from on an
individual machine, we also need to know which machine is best to
target a job onto within a large cluster;

- Low, predictable CPU usage;

- Among cold pages, differentiate between these that are immediately
reclaimable and these that would require a disk write.

We use a very simple approach, scanning memory at a fixed rate and
identifying pages that haven't been touched in a number of scans. We
are currently switching from a fakenuma based implementation (which we
don't think is very upstreamable) to a memcg based one. We think this
could be of interest to the wider community & would like to discuss
requirement with other interested folks.

-- 
Michel "Walken" Lespinasse
A program is never fully debugged until the last user dies.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: [LSF/MM TOPIC][ATTEND]cold page tracking / working set estimation

[KOSAKI Motohiro]

> Google uses an automated system to assign compute jobs to individual
> machines within a cluster. In order to improve memory utilization in
> the cluster, this system collects memory utilization statistics for
> each cgroup on each machine. The following properties are desired for
> the working set estimation mechanism:
> 
> - Low impact on the normal MM algorithms - we don't want to stress the
> VM just by enabling working set estimation;
> 
> - Collected statistics should be comparable across multiple machines -
> we don't just want to know which cgroup to reclaim from on an
> individual machine, we also need to know which machine is best to
> target a job onto within a large cluster;
> 
> - Low, predictable CPU usage;
> 
> - Among cold pages, differentiate between these that are immediately
> reclaimable and these that would require a disk write.
> 
> We use a very simple approach, scanning memory at a fixed rate and
> identifying pages that haven't been touched in a number of scans. We
> are currently switching from a fakenuma based implementation (which we
> don't think is very upstreamable) to a memcg based one. We think this
> could be of interest to the wider community & would like to discuss
> requirement with other interested folks.

Hi, Michel and Program comittees

I'm not sure what status is MM track. but if this proposal was accepted and
seat number is allowed, I'd like to attend and discuss the issue with other
cloud intersted developers.

Thanks.


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>