[Gluster-devel] Disk power-saving friendliness

[Gordan Bobic]

On Wed, 01 Jul 2009 12:32:20 +0530, Shehjar Tikoo <shehjart at gluster.com>
wrote:

> Furthermore, all writes are forced to disk at a close().

Hmm, that could be it. But would this not be caught by the usual write
delaying functionality in laptop_mode?

>> I find that if I put the server into "laptop mode" to conserve power 
>> by spinning down the disk writes still tend to wake them up.
> 
> Are you sure they're writes and not reads?

Anything is possible, but I have lots of RAM in this server and have 4MB of
read-ahead set, followed by head -c1 on every file to pre-buffer
everything. In that state I wouldn't expect the disks to ever spin up
except to write out the data every 10 minutes.

> And are you sure it is
> GlusterFS process that is issuing requests that need service from disk?

I'm not 100% sure at the moment. It could also be gluster's logging since
it's going raw rather than through syslog.

>> Are the any particular paramters/options that should be used in this 
>> mode to minimize disk wake-ups? Disable direct I/O? Does write-behind
>>  work for local disks? Any suggestions on this?
> 
> 1. Try spinning down without GlusterFS running. See if the disk still
> wakes-up. It is possible a different application wakes up the disk.
> 
> 2. With glusterfsd running, try unmounting all the mount points to
> this server and leave the server running. It might help us pin-point
> if the server is issuing disk requests without us knowing about it.
> 
> 3. Keep GlusterFS server running along with all the mount points
> mounted. But do not have any applications running on the mount points.
> This will help us determine if the GlusterFS client side stack is
> issuing disk requests even though no application requests are coming
> through.

OK, I'll try all that and report back. I'll also apply the patch I posted
here:
https://bugzilla.redhat.com/show_bug.cgi?id=223722
and see if that helps.

Gordan