Re: i/o bandwidth controller infrastructure

[Andrea Righi]

Divyesh Shah wrote:
> > This is the core io-throttle kernel infrastructure. It creates the
> > basic
> > interfaces to cgroups and implements the I/O measurement and
> > throttling
> > functions.
>
> I am not sure if throttling an application's cpu usage by explicitly  
> putting it to sleep
> in order to restrain it from making more IO requests is the way to go  
> here (though I can't think
> of anything better right now).
> With this bandwidth controller, a cpu-intensive job which otherwise  
> does not care about its IO
> performance needs to be pin-point accurate about IO bandwidth  
> required in order to not suffer
> from cpu-throttling. IMHO, if a cgroup is exceeding its limit for a  
> given resource, the throttling
> should be done _only_ for that resource.
>
> -Divyesh

Divyesh,

I understand your point of view. It would be nice if we could just
"disable" the i/o for a cgroup that exceeds its limit, instead of
scheduling some sleep()s, so the tasks running in this cgroup would be
able to continue their non-i/o operations as usual.

However, how to do if the tasks continue to perform i/o ops under this
condition? we could just cache the i/o in memory and at the same time
reduce the i/o priority of those tasks' requests, but this would require
a lot of memory, more space in the page cache, and probably could lead
to potential OOM conditions. A safer approach IMHO is to force the tasks
to wait synchronously on each operation that directly or indirectly
generates i/o. The last one is the solution implemented by this
bandwidth controller.

We could collect additional statistics, or implement some heuristics to
predict the tasks' i/o patterns in order to not penalize cpu-bound jobs
too much, but the basic concept is the same.

Anyway, I agree there must be a better solution, but this is the best
I've found right now... nice ideas are welcome.

-Andrea
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/