Skip site navigation (1)Skip section navigation (2)
Date:      22 Aug 1998 19:49:39 GMT
From:      freebsd-net-list@salford.ac.uk
To:        freebsd-net@FreeBSD.ORG
Subject:   Re: 2.2.6 net performance and panic with 1000's of sockets open
Message-ID:  <19980822195713.11147.qmail@ananke.salford.ac.uk>

next in thread | raw e-mail | index | archive | help
Mark Powell (mark@nospam.salford.ac.uk) wrote:
: In article <6pl4lk$a51$3@ocean.cup.hp.com>, Rick Jones <foo@bar.baz> wrote:
: >I think Steinar meant ttcp the benchmark, not T/TCP the TCP protocol
: >extensions. As for netperf only getting 64 Mbit/s, it would help if
: >you could do a cut and paste of your command lines so we can see the

Of course, I was silly and failed to ask for the netperf client output.

: If I use:

: $ route -n change <server> -recvpipe 65536 -sendpipe 65536

Did the socket buffer size reports change in the netperf output when
you did that route command?

: On the client, I've seen 73Mbit/s.
: If I do a UDP_STREAM I get 95.8Mbit/s.

One has to be *very* careful with the UDP_STREAM test and really only
report the recevie rate. There are some systems and/or drivers out
there which do not report ENOBUFs when the transmit queues overflow.

[you may have actually reported the receive rate, I am just paranoid
 when it comes to the UDP_STREAM test...I very nearly did not include
 it in netperf ]

If you were indeed getting 95 Mbit/s with UDP_STREAM, it is possible
that the window size, even at 65535 bytes was not enought to deal with
the latency of the connection.

One first-order appoximation of the limit of performance with TCP is
W/RTT where W is the window size, and RTT is the round-trip time. Some
NICs have higher latency than others...there have been some NICs with
really bad latency, which is one reason why there is  a TCP_RR test in
netperf.

If there is high latency (high is a relative term) increasing the
window can help.

It is also a good idea to check the netstat stats for retransmitted
segments. The occasional loss of a datagram will not show-up in the
UDP_STREAM results, but it can in a TCP_STREAM test.

rick jones
--
these opinions are mine, all mine; HP might not want them anyway... :)
feel free to email, or post, but please do not do both...
my email address is raj in the cup.hp.com domain...

To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-net" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?19980822195713.11147.qmail>