Performance drop to very near 0 when memcachetest runs on the cluster with more than one node

When the memcachetest runs on the cluster with more than one node, ops/sec dropped to almost 0 after 2-3 minutes and stayed around there for the length of the process.

So, I restarted the memcachetest again, and ops/sec was going up, but dropped to 0 again after 2-3 minutes.

Steve and I ran the python client and found that if a request from the python client is directly sent to the memcache process, it sends a response back to the client immediately. However, if a request is routed through the moxi, the response time was very slow.