The following errors can be seen in the segment logs:
interconnect encountered a network error, please check your network
Failed to send packet (seq 34) to <IP address>:<port>
The issue is caused mainly by network limitations. Increasing the network queues/buffers and/or reducing the network load can help avoiding the errors.
1. Use the interconnect proxy with the instructions in Configuring Proxies for the Greenplum Interconnect. Using the interconnect proxy can reduce the number of connections between the hosts and the number of ports used by Greenplum.
2. Increase the GUC gp_interconnect_tcp_listener_backlog to 512
# Check the current value. Default for 6.x is 128 and default for 7.x is 256
gpconfig -s gp_interconnect_tcp_listener_backlog
# Change the value and reload the configuration files
gpconfig -c gp_interconnect_tcp_listener_backlog -v 512
gpstop -u
The GUC "gp_interconnect_tcp_listener_backlog" is not documented. The values can be 0-65535.
3. Check the value of kernel parameter "net.core.somaxconn"
sysctl net.core.somaxconn
The value can be increased to 10000. See Tuning the TCP listen backlog to process a high number of TCP connection attempts for information on the parameter.
Other KBs related to the error messages:
Greenplum query fails with could not read() message from seqserver
Mirror segments are continually failing or "interconnect encountered a network error"