This problem is related with having a lot of TIME_WAIT
connections:
netstat -an | grep TIME_WAIT | wc -l > 20000
A solution is to tune your kernel Linux up inputting the following parameters:
echo 32000> /proc/sys/net/ipv4/tcp_max_syn_backlog echo 30> /proc/sys/net/ipv4/tcp_fin_timeout
This is not meant to be used with load balancing/clustering environments:
echo 1 > /proc/sys/net/ipv4/tcp_tw_recycle