[Gluster-users] Very slow directory listing and high CPU usage on replicated volume

[pkoelle]

Hi,

from Joe's findings it seems the IO subsystem has little impact on 
performance. If a ramdisk doesn't help with small files and directory 
listings the problem is probably gluster itself.

That being said IMO this is a general limitation in the design. There is 
no central lock manager and last time I checked gluster had to stat() on 
each brick to find the most recent entry.
FUSE overhead and context switches make things worse, but to what extent 
and which part is responsible for high CPU/high latency is still 
unknown. AFAIK noone has measured in detail or the findings where such 
that internal policy is not to publish the results.

IIRC there was/is a stat cache translator but that is just a workaround 
and will lead to stale data.

cheers
  Paul

Am 06.11.2012 10:35, schrieb Fernando Frediani (Qube):
> Joe,
>
> I don't think we have to accept this as this is not acceptable thing. I have seen countless people complaining about this problem for a while and seems no improvements have been done.
> The thing about the ramdisk although might help, looks more a chewing gun. I have seen other distributed filesystems that don't suffer for the same problem, so why Gluster have to ?
>
> -----Original Message-----
> From: gluster-users-bounces at gluster.org [mailto:gluster-users-bounces at gluster.org] On Behalf Of Joe Landman
> Sent: 05 November 2012 15:07
> To: gluster-users at gluster.org
> Subject: Re: [Gluster-users] Very slow directory listing and high CPU usage on replicated volume
>
> On 11/05/2012 09:57 AM, harry mangalam wrote:
>> Jeff Darcy wrote a nice piece in his hekafs blog about 'the importance
>> of keeping things sequential' which is essentially about the
>> contention for heads between data io and journal io.
>> <http://hekafs.org/index.php/2012/11/the-importance-of-staying-sequent
>> ial/> (also congrats on the Linux Journal article on the glupy
>> python/gluster approach).
>>
>> We've been experimenting with SSDs on ZFS (using the SSDs fo the ZIL
>> (journal)) and while it's provided a little bit of a boost, it has not
>> been dramatic.  Ditto XFS.  However, we did not stress it at all with
>> heavy loads
>
> An issue you have to worry about is if the SSD streaming read/write path is around the same speed as the spinning rust performance.  If so, this design would be a wash at best.
>
> Also, if this is under Linux, the ZFS pathways may not be terribly well optimized.
>
>> in a gluster env and I'm now thinking that there is where you would
>> see the improvement. (see Jeff's graph about how the diff in
>> threads/load affects IOPS).
>>
>> Is anyone running a gluster system with the underlying XFS writing the
>> journal to SSDs?  If so, any improvement?  I would have expected to
>> hear about this as a recommended architecture for gluster if it had
>> performed MUCH better, but
>
> Yes, we've done this, and do this on occasion.  No, there's no dramatic speed boost for most use cases.
>
> Unfortunately, heavy metadata ops on GlusterFS are going to be slow, and we simply have to accept that for the near term.  This appears to be independent of the particular file system, or even storage technology.
> If you aren't doing metadata heavy ops, then you should be in good shape.  It appears that mirroring magnifies the metadata heavy ops significantly.
>
> For laughs, about a year ago, we set up large ram disks (tmpfs) in a cluster, put a loopback device on them, then a file system, then GlusterFS atop this.  Should have been very fast for metadata ops.  But it wasn't.  Gave some improvement, but not significant enough that we'd recommend doing "heroic" designs like this.
>
> If your workloads are metadata heavy, we'd recommend local IO, and if you are mostly small IO, an SSD.
>
>
>
>


-- 
https://ssl.facebook.com/help/contact.php?show_form=delete_account