One of our customers asked if Vertica has any Network / TCP / UDP Tuning recommendations such as:
net.core.wmem_max
net.core.rmem_max
net.core.wmem_default
net.core.rmem_default
net.ipv4.tcp_rmem
net.ipv4.tcp_wmem
Does the following can be applied to any large Vertica cluster?
Please share the conditions one should consider to match the calculations to a given system.
The following found in a Confluence document from 2013 by Ben Vandiver
http://confluence.verticacorp.com:8080/download/attachments/29789547/vertica-network-architecture-crane.pptx?api=v2
OS parameters – example changes for 10Gb network
# default non-tcp buffer size
net.core.rmem_default = 262144
net.core.wmem_default = 262144
# maximum non-tcp buffer size
net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
# large backlog in kernel before processing into tcp stack
net.core.netdev_max_backlog = 100000
# lowest, default, maximum tcp buffers
net.ipv4.tcp_rmem = 8192 262144 8388608
net.ipv4.tcp_wmem = 8192 262144 8388608
# minimum udp socket size
net.ipv4.udp_rmem_min = 16384
net.ipv4.udp_wmem_min = 16384
# allow sockets to take up the unrealistic maximum of 64gb before we start restricting memory
net.ipv4.tcp_mem = 16777216 16777216 16777216
net.ipv4.udp_mem = 16777216 16777216 16777216
# socket open
net.core.somaxconn = 1024
