I’m trying to sysbench a Percona XtraDB cluster but I’m finding that if I run the test on a single node (started with wsrep_provider=‘none’) the performance is good, after that I stop the node and start the cluster with
service start bootstrap-pxc
Then the performance goes down a little, but it is still good.
Now, we start a second node in another server with the same hardware located in the same network, connected by two Gb Ethernet adapters. It makes SST, and both nodes get Synced.
When I run the test against the first node, the performace goes down a lot.
I thought it is a flow control issue, so to confirm, I relaxed flow control starting second node with “gcs.fc_limit=1G; gcs.fc_factor=1.0”, and launch the test again.
In this test the performance is good again, but I can see that the recv_queue in the second node grows very fast.
So, my conclusion is that second node is slow when appliying the writesets, but if I change the servers roles I find the same situation, the “slave” node is allwais slower than the master.
I can’t imagine why…