[ietf78-tech] linux interface/bridge tuning historical notes...

[kemp]

For future reference, here are some of the OS settings
we have typically messed with when we did bridges in the past.
We didn't mess with these on the IETF nacs for lack of time.

Other things to always check are driver version,
and neighbor cache behavior (ARP).  We would typically set
Intel drivers to use the 4096 rx and tx buffer
on the card:

	ethtool -G eth1 rx 4096 tx 4096
	ethtool -G eth2 rx 4096 tx 4096

Running a traffic generator through the box and testing before/after
settings is ideal when you do stuff like this...

/jgk


# 4KB send buffer, 20,480 connections max at worst case
echo 83886080  > /proc/sys/net/core/wmem_max
echo 83886080  > /proc/sys/net/core/wmem_default

# 16KB receive buffer, 20,480 connections max at worst case
echo 335544320 > /proc/sys/net/core/rmem_max
echo 335544320 > /proc/sys/net/core/rmem_default

# Max open files
echo 65536 > /proc/sys/fs/file­max

# Fast port recycling (TIME_WAIT)
echo 1 >/proc/sys/net/ipv4/tcp_tw_recycle
echo 1 >/proc/sys/net/ipv4/tcp_tw_reuse

# TIME_WAIT buckets increased
echo 65536 > /proc/sys/net/ipv4/tcp_max_tw_buckets

# FIN timeout decreased
echo 15 > /proc/sys/net/ipv4/tcp_fin_timeout

# SYN backlog increased
echo 65536 > /proc/sys/net/ipv4/tcp_max_syn_backlog

# SYN cookies enabled
echo 1 > /proc/sys/net/ipv4/tcp_syncookies

# Local port range maximized
echo "1024 65535" > /proc/sys/net/ipv4/ip_local_port_range

# Netdev backlog increased
echo 100000 > /proc/sys/net/core/netdev_max_backlog

# Interface transmit queuelen increased
ifconfig eth0 txqueuelen 10000


-- 
John Kemp
kemp at network-services.uoregon.edu
RouteViews Engineer
541-346-1714