[Gluster-devel] CPU usage in AFR

[Gordan Bobic]

Vikas Gorur wrote:
> Gordan Bobic wrote:
>> Hi,
>>
>> I'm noticing something that could be considered weird. I have a 3-node 
>> rootfs-on-gluster cluster, and I'm doing a big yum update on one of 
>> the nodes. On that node and one of the other nodes, glusterfsd is 
>> using about 10-15% CPU. But on the 3rd node, glusterfs is using < 0.3% 
>> CPU.
>>
>> I'm guessing this has something to do with the way locking backup and 
>> fail-over is implemented (IIRC there is a master and a backup 
>> lock/state server, rather than replicating this information to all the 
>> servers). Can anyone hazard a guess and confirm or deny the possible 
>> cause of this discrepancy?
 >
> Hm, this is indeed a little strange. The first subvolume is used as the 
> lock server, so that could possibly explain why one of the servers have 
> higher CPU usage. However, this doesn't explain why *two* of the servers
> have high CPU usage.
> 
> Are you sure it is not the case that the *client* on one node (I'm assuming
> all your three nodes act as servers and clients) and the server on 
> another node are using CPU?

That is plausible since I am using single-process client and server. Is 
there a way to tell on a running glfs cluster which node is the current 
lock master? The process creating the load was running on the first 
listed volume, so I would have expected this to be the primary lock server.

Gordan