Anyway to improve Gearmand speed?

[Chris Goffinet]

I've been doing some tests lately on gearmand using Perl and PHP. I  
noticed that for Linux, to get any acceptable speeds TCP_NODELAY must  
be enabled because of the small packet sizes of the protocol. I  
noticed that if you setup 1 worker in perl (task that just returned  
true), and 1 client to call that task 25 times, on average would take  
~50ms (same host). This isn't always true on BSD platform which can be  
usually ~20ms. Any suggestions that I could poke into to potentially  
improve this?

The type of server used is a PE860,  1 x Xeon 3050 2.13GHz, 4GB, RHEL  
4.4

-Chris