I've got a small setup, one machine that acts as a load balancer and two machines that do all the work. The load balancer runs nginx (static content + php proxying to workers) and mysql, the two workers run php5-fpm and memcached (pooled between workers).

Here's beancounters for the balancer:

tcpsndbuf         2171848      2386280     10000000     20000000      3947733
tcprcvbuf         1248288      1669504     10000000     20000000            0

Here's worker 1:

tcpsndbuf          951976      1262672     20000000     40000000            0
tcprcvbuf          278528       393496     20000000     40000000            0

Here's worker 2:

tcpsndbuf          989888       527472     20000000     40000000            0
tcprcvbuf          212992       452520     20000000     40000000            0

The balancer has 1GB ram, the two workers have 2GB ram each.

What is eating my send buffer?