[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Linux-cluster] Slowness above 500 RRDs



David Teigland <teigland redhat com> writes:

> On Mon, May 21, 2007 at 04:42:11PM +0200, Ferenc Wagner wrote:
>
>> One more thing: the old infrastructure became fast when only one (or
>> sometimes two) node kept the shared volume mounted.  Now this effect
>> is lost as well, a single node or all three nodes mounting the volume
>> shows the same not-so-stellar performance.  Is this expected, or do I
>> make a mistake somewhere?
>
> That's expected for plocks, but not flocks.

Just checked again, it doesn't make any difference with either.  The
network traffic is no different, on node 1 I can see packets leaving
for 2 and arriving from 3.  This of course varies with the number of
nodes participating the cluster, eg. leaving on node 3 drops it from
the cycle, leaving also on node 2 results in multicast storm only.
But it doesn't make any difference in performance, turning to the
network is too slow in any case.  Still I wonder why it's necessary
when nobody else mounts the volume -- not to mention the multicast
traffic when the node is the only member of the cluster.

Counting master locks (as you described in your other mail) will
hopefully shed some light on this issue.  I'm working on it, but some
kernel stability issues got in the way.
-- 
Thanks,
Feri.


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]