[Linux-cluster] GFS cluster / DLM locking - Mostly idle but high load

[Gordan Bobic]

On Wed, 17 Oct 2007, Marc Grimme wrote:

>>>> I have a cluster (3 nodes at the moment, may grow up to 16) for handling
>>>> a lot of small files (Maildir). When I test the system by sending around
>>>> 3-5 messages/second I see the load on the cluster nodes go up to about
>>>> 20-30, even though the CPUs on the cluster are about 90% idle at all
>>>> times.
>>>>
>>>> I am guessing that this is due to the clustered machines waiting for DLM
>>>> locks to be established, which causes a lot of processes to be fighting
>>>> to run, but since they don't get to run very soon, they back up and
>>>> cause the load averages to go up.
>>>>
>>>> Assuming the DLM runs over the interface specified by IP and MAC in
>>>> cluster.conf, it is running over gigabit ethernet.
>>>>
>>>> Are there any configuration changes or tuning parameters I can apply to
>>>> DLM to alleviate this condition? The machine I'm running the test from
>>>> (the one sending messages) is about 1/4 of the spec of each of the
>>>> cluster nodes, and it's running a load average of about 0.4. It seems
>>>> crazy that a single low-spec node should be able to completely overwhelm
>>>> a cluster 12x it's spec several times over.
>>>
>>> I don't know alot about GFS but since no one else has replied yet, my
>>> understanding is that it's not suitable for an applications like what
>>> you describe (many small files being opened frequently). I think GFS2,
>>> which is still a tech preview, has been redesigned to improve this
>>> situation.
>>
>> Indeed, I am aware that GFS2 is still broken, but I seem to be getting no
>> worse a performance out of GFS than I get out of NFS. The only penalty is
>> the high load, but the throughput is actually similar. The advantage that
>> makes GFS win is that I don't need an arbitrating server to handle the NFS
>> exports, which makes the clustering and redundancy a bit tidier.
>
> with your testing did you also try to adapt the size of the
> rsbtbl_size/lkbtbl_size? I would be quite interested if this increases your
> performance or not.

I cannot find these files in /proc (that's where they are implied to be in 
the docs). Can you please point me in the right direction?

> Do you have lot of small files?

Yes. The problem doesn't seem to be so bad when files are in different 
directories, but when lots of files are being written to the same 
directory, the load goes up quite badly.

Gordan