[Gluster-devel] Possible Performance Tweak for unfsd Users

[Gordan Bobic]

Shehjar Tikoo wrote:
> Gordan Bobic wrote:
>> I can't figure out why this might be the case, but it would appear
>> that when unfsd is bound to a custom port and not registered with
>> portmap, the performance is massively improved.
>>
>> I changed my init.d/unfsd script as follows, in the start option:
>>
>> - /usr/sbin/unfsd -i ${pidfile} + /usr/sbin/unfsd -e /etc/uexports
>> -i ${pidfile} -m 12049 -n 12049 -p
> 
> How was your experience with LD_PRELOADed booster instead of going
> through a FUSE mountpoint?

I never tried it. I was informed by one of the developers that booster 
is not particularly helpful, can cause stability problems and is soon to 
be deprecated in the 3.0.x branch.

>> cat /etc/uexports /home           10.2.0.0/16(rw,insecure)
>>
>> On the client side I am mounting with:
>>
>> server:/home  /home  nfs 
>> defaults,nolock,hard,noatime,proto=udp,rsize=32768,wsize=32768,port=12049,mountport=12049 
>>
>>  0 0
>>
> 
> You're using UDP. Are you sure NFS client did not default to TCP
> earlier when you were not specifying non-default ports?

Considering this is on a LAN, I would expect the difference in 
performance between TCP and UDP to be approximately the difference 
implied by the header sizes: 8 bytes for UDP vs. 20 bytes for TCP. I 
wouldn't imagine that to make that much difference on the payload of 32KB.

I can try TCP vs. UDP but I'm reasonably sure I've always mounted NFS 
volumes via UDP. In fact, unfsd shares refuse to mount TCP at the moment.

>> All of the horrible latency and laggyness is completely gone! Could
>> it be that it is unfsd's interraction with rpc/portmap/mountd that
>> is the cause of a lot of the performance issues?
> 
> unfsd only needs to register with portmap once when it starts so I
> dont think that is the reason.

I didn't think it would be, either, but something is quite clearly 
different and is making a massive difference.

Gordan