RE: SUMMARY:RE: TCP_keepalive

Apologies , fat finger does it again.  Here are Eric and Sam's replies

Eric's:

We had to tweak around with the keepalive here.

Typically, short keepalive values are good for web servers in example (lots 
of short and unreliable connections). Obvisouly long keepalive are good for 
long term connections such as applications having a connection to a
database...

We had to shorten this value and the "time wait" one because we had lots of 
open unused client sockets which were left open, consuming resources.

But there is something: the keep alive parameter on the server will only be 
taken into account if  "keep alive" is specified in the options when 
creating a TCP socket. For example, it will have no effect if a client 
application doesn't explicitely set the keep alive flag when connecting to 
a server...

You might want to have a look at the "tcp_time_wait_interval" parameter as 
well...

Sam's :


I'm a sysadmin at an ISP in the UK - as such I deal with many busy 
machines in a production environment. In my experience you don't want to 
go lower than 60000ms (even on an extremely busy machine) with 
tcp_keepalive. I usually dump into /etc/rc2.d in this simple script on 
my boxes as a baseline to work with:

#!/bin/sh
# --------------------------------------------------------------- #
#
# S70stacktune 0.13
# 23/09/2004 Sam Nelson <sam@unix.ms>
# Script to tune and sanitise the TCP stack on busy Solaris boxes.
#

UREVISION=`uname -r`

fill () {
awk '{leninput=length($($NF)); fill=63-leninput ; for (i=1; i< fill; 
i++) fillchar=fillchar"." ; printf $($NF) fillchar}'
}

setparams () {
VAL=`/usr/sbin/ndd -set $1 $2 $3`
printf "Value of $1 $2 is: " | fill
printf " ${VAL} ($3)\n"
}


# Decrease the tcp time wait interval
if [ ${UREVISION} = "5.6" ]; then
printf "Decrease the tcp time wait interval\n"
   setparams /dev/tcp tcp_close_wait_interval 60000
else
printf "Decrease the tcp time wait interval\n"
   setparams /dev/tcp tcp_time_wait_interval 60000
fi

# Speed up the flushing of half-closed connection in state FIN_WAIT_2
printf "Flushing of half-closed connection in state FIN_WAIT_2\n"
setparams /dev/tcp tcp_fin_wait_2_flush_interval 67500ms

# Increase the receive and transmit window sizes
printf "Increase the receive and transmit window sizes\n"
setparams /dev/tcp tcp_xmit_hiwat 65535
setparams /dev/tcp tcp_recv_hiwat 65535

# decrease the retransmit interval
printf "Decrease the retransmit interval\n"
setparams /dev/tcp tcp_rexmit_interval_max 60000ms

# increase number of half-open connections
printf "Increase number of half-open connections\n"
setparams /dev/tcp tcp_conn_req_max_q0 4096

# increase number of simultaneous connections
printf "Increase number of simultaneous connections\n"
setparams /dev/tcp tcp_conn_req_max_q 1024

# Decrease TCP connection abort interval
printf "Decrease TCP connection abort interval\n"
setparams /dev/tcp tcp_ip_abort_interval 60000

# Decrease TCP Keepalive Interval
printf "Decrease TCP Keepalive Interval\n"
setparams /dev/tcp tcp_keepalive_interval 60000

printf "Ok, tuning complete\n"
-------------------------------------------------------

This usually does the trick.

Hope this helps, and please summarise!

Regards

Sam

-----Original Message-----
From: Rich Bonfoey 
Sent: Friday, April 01, 2005 11:12 AM
To: Rich Bonfoey; 'Sun Managers (sunmanagers@sunmanagers.org)'
Subject: SUMMARY:RE: TCP_keepalive

Many thanks to Eric Voisard and Sam Nelson for sharing their expertise.
Eric provided
insight to another variable tcp_time_wait_interval, that it can have an
affect on connections.  Sam , very graciously  provided a script that he
uses to tune and sanitize the tcp stack

Below is the original message
With Eric and Sam's replays

Thanks again for the info

Rich



-----Original Message-----
From: Rich Bonfoey 
Sent: Thursday, March 31, 2005 9:55 AM
To: Sun Managers (sunmanagers@sunmanagers.org)
Subject: TCP_keepalive

Greetings

Have any of you, had much dealings with the tcp_keepalive ?

On Solaris 8 and 9 the value is 7200000 or 2 hours.  What are  some of the
pit falls if this value was lowered to about 10 minutes.  I have read the
RFC but was looking for people who have had real experience with tweaking
this value.

Part of the reason for tweaking, is we have a large number of PC and OS 10
users connecting to several different systems and they have been having
performance issues.  Our dba decided to bring in a consultant ( that's
another story ) and they said this value should be 5 minutes.  I am not
quite buying it, as I have read that there are multiple opinion on the
correct use of tcp_keepalive.



Richard Bonfoey
The News Tribune
Information Systems
Successfully Meeting the Business Needs of
The News Tribune through Information Technology
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers