[Gluster-users] Self-heal and high load

[Toby Corkindale]

On 11/05/13 00:40, Matthew Day wrote:
> Hi all,
>
> I'm pretty new to Gluster, and the company I work for uses it for
> storage across 2 data centres. An issue has cropped up fairly recently
> with regards to the self-heal mechanism.
>
> Occasionally the connection between these 2 Gluster servers breaks or
> drops momentarily. Due to the nature of the business it's highly likely
> that files have been written during this time. When the self-heal daemon
> runs it notices a discrepancy and gets the volume up to date. The
> problem we've been seeing is that this appears to cause the CPU load to
> increase massively on both servers whilst the healing process takes place.
>
> After trying to find out if there were any persistent network issues I
> tried recreating this on a test system and can now re-produce at will.
> Our test system set up is made up of 3 VMs, 2 Gluster servers and a
> client. The process to cause this was:
> Add in an iptables rule to block one of the Gluster servers from being
> reached by the other server and the client.
> Create some random files on the client.
> Flush the iptables rules out so the server is reachable again.
> Force a self heal to run.
> Watch as the load on the Gluster servers goes bananas.
>
> The problem with this is that whilst the self-heal happens one the
> gluster servers will be inaccessible from the client, meaning no files
> can be read or written, causing problems for our users.
>
> I've been searching for a solution, or at least someone else who has
> been having the same problem and not found anything. I don't know if
> this is a bug or config issue (see below for config details). I've tried
> a variety of different options but none of them have had any effect.


For what it's worth.. I get this same behaviour, and our gluster servers 
aren't even in separate data centres. It's not always the self-heal 
daemon that triggers it -- sometimes the client gets in first.

Either way -- while recovery occurs, the available i/o to clients drops 
to effectively nothing, and they stall until recovery completes.

I believe this problem is most visible when your architecture contains a 
lot of small files per directory. If you can change your filesystem 
layout to avoid this, then you may not be hit as hard.
(eg. Take an MD5 hash of the path and filename, then store the file 
under a subdirectory named after the first few characters in the hash. 
(2 chars will divide the files-per-directory by ~1300, three by ~47k) 
eg. "folder/file.dat" becomes "66/folder/file.dat")


I've given up on GlusterFS though; have a look at Ceph and RiakCS if 
your systems suit Swift/S3 style storage.

-Toby