Hi Vivek,
I am trying to optimize user-experience on a busy NFS server.
I think much could be achieved if the following was true.
get a response to file system meta data operations (opendir,
readdir, stat, mkdir, unlink) within 200 ms even under heavy
read/write strain ...
In the course of my research I also tried the io-controller patches.
My test setup consists of several tar processes keeping a disk busy
by packing and unpacking Linux kernels.
I was able to bring read and write bandwidth into balance by
putting the reading and writing tars in to different cgroups.
Unfortunately this did not seem to help my goal since meta data
operations do not seem to get treated differently from normal
operations (or maybe even worse?)
Is there a way to get io-controller to help me with this?
cheers
tobi
--
Tobi Oetiker, OETIKER+PARTNER AG, Aarweg 15 CH-4600 Olten, Switzerland
http://it.oetiker.ch [email protected] ++41 62 775 9902 / sb: -9900
On Wed, Sep 16, 2009 at 02:58:52PM +0200, Tobias Oetiker wrote:
> Hi Vivek,
>
> I am trying to optimize user-experience on a busy NFS server.
>
> I think much could be achieved if the following was true.
>
> get a response to file system meta data operations (opendir,
> readdir, stat, mkdir, unlink) within 200 ms even under heavy
> read/write strain ...
>
> In the course of my research I also tried the io-controller patches.
>
> My test setup consists of several tar processes keeping a disk busy
> by packing and unpacking Linux kernels.
>
> I was able to bring read and write bandwidth into balance by
> putting the reading and writing tars in to different cgroups.
>
> Unfortunately this did not seem to help my goal since meta data
> operations do not seem to get treated differently from normal
> operations (or maybe even worse?)
>
> Is there a way to get io-controller to help me with this?
Hi tobi,
Is it better with vanilla CFQ (without io controller). I see that CFQ
preempts the ongoing process if it receives a meta data request and that
way it provides faster response.
If yes, then similar thing should work for IO controller also. Wait there
is one issue though. If a file system request gets backlogged in a group
while a different group was being served, then preemption will not happen
and that's probably the reason you are not seeing better latencies.
I think there are two ways to handle this in IO controller.
- Put the meta data requesting processes at the front of the service tree
in respective group. This will make sure that even if there are other
sequential readers or heavy writers in the group, this request gets
served quickly.
I will write a small patch for this. I think that should help you.
Thanks
Vivek
Hi Vivek,
Today Vivek Goyal wrote:
> On Wed, Sep 16, 2009 at 02:58:52PM +0200, Tobias Oetiker wrote:
> > Hi Vivek,
> >
> > I am trying to optimize user-experience on a busy NFS server.
> >
> > I think much could be achieved if the following was true.
> >
> > get a response to file system meta data operations (opendir,
> > readdir, stat, mkdir, unlink) within 200 ms even under heavy
> > read/write strain ...
> >
>
> Hi tobi,
>
> Is it better with vanilla CFQ (without io controller). I see that CFQ
> preempts the ongoing process if it receives a meta data request and that
> way it provides faster response.
>
> If yes, then similar thing should work for IO controller also. Wait there
> is one issue though. If a file system request gets backlogged in a group
> while a different group was being served, then preemption will not happen
> and that's probably the reason you are not seeing better latencies.
I only looked at io-controller once I did not get ahead with
vanilla cfq ( and the other schedulers for that matter).
I am working on a system with and Areca HW Raid6 with battery buffered cache.
I assume that if linux manages to fill the cache of the hwraid with
write requests it will eventually flush them and maybe not react
well to competing read requests at that level as well (just
guessing) I have disabled the cache, to see if this has a positive effect, but
the only gain was, that the everything became slower. Metadata
operations still were painfuly slow.
> I think there are two ways to handle this in IO controller.
> - Put the meta data requesting processes at the front of the service tree
> in respective group. This will make sure that even if there are other
> sequential readers or heavy writers in the group, this request gets
> served quickly.
in my test setup this may be possible, but IRL I am dealing with an
NFS server, so I can not deal with individual processes (they are
on the clients and not known on the nfs sever I guess).
I am a bit at a loss as to how I should best configure
io-controller for such a situation since it seems to relie on
process ids for all its work ...
maybe fairness between different clients and userids might be
interesting here ... can cgroups deal with this?
> I will write a small patch for this. I think that should help you.
cool ..
cheers
tobi
> Thanks
> Vivek
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
>
--
Tobi Oetiker, OETIKER+PARTNER AG, Aarweg 15 CH-4600 Olten, Switzerland
http://it.oetiker.ch [email protected] ++41 62 775 9902 / sb: -9900
On Wed, Sep 16, 2009 at 05:03:11PM +0200, Tobias Oetiker wrote:
> Hi Vivek,
>
> Today Vivek Goyal wrote:
>
> > On Wed, Sep 16, 2009 at 02:58:52PM +0200, Tobias Oetiker wrote:
> > > Hi Vivek,
> > >
> > > I am trying to optimize user-experience on a busy NFS server.
> > >
> > > I think much could be achieved if the following was true.
> > >
> > > get a response to file system meta data operations (opendir,
> > > readdir, stat, mkdir, unlink) within 200 ms even under heavy
> > > read/write strain ...
> > >
> >
> > Hi tobi,
> >
> > Is it better with vanilla CFQ (without io controller). I see that CFQ
> > preempts the ongoing process if it receives a meta data request and that
> > way it provides faster response.
> >
> > If yes, then similar thing should work for IO controller also. Wait there
> > is one issue though. If a file system request gets backlogged in a group
> > while a different group was being served, then preemption will not happen
> > and that's probably the reason you are not seeing better latencies.
>
> I only looked at io-controller once I did not get ahead with
> vanilla cfq ( and the other schedulers for that matter).
Thinking more about it. You need to first identify where the latency is
coming from. io controller can help you in providing better latencies if
latencies are coming from other processes competing for the disk and delay
is coming from the fact that metadata request from nfs thread is blocked
behind other requests.
I see that CFQ preempts a current queue if it is not doing metadata
operation. That means CFQ is providing immediate access to disk(apart from
requests in dispatch queue or in hardware).
So I am not even sure where the delay is coming from. Once we figure that
out, I can tell whether io controller will help you or not. In general, io
controller will be able to help you if delay is coming from the IO
scheduler because you are queued behind other requests. But in case of
CFQ, that does not seem to be the case.
Can you do some blktraces and figure out where the delay is coming from?
Thanks
Vivek