Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!tut.cis.ohio-state.edu!ucbvax!fstohp.crd.ge.com!keith
From: keith@fstohp.crd.ge.com (Keith D Gregory)
Newsgroups: comp.protocols.tcp-ip
Subject: (none)
Message-ID: <8905151447.AA24637@crdgw1.ge.com>
Date: 15 May 89 14:48:51 GMT
Sender: daemon@ucbvax.BERKELEY.EDU
Organization: The Internet
Lines: 42

Subject: Re: UDP Connections on LOOP interface

joshua@atherton.com (Joshua Levy) writes:
>What about the relative speeds?  Do you have any idea how fast msgsnd
>is compared to socket(127.0.0.1)?  If so, I'd like to know.  Thanks.

The raw numbers I generated (all times in milliseconds):

   Method: UDP on 127.0.0.1
      Number of packets: 1000
      Maximum time:        39.956
      Minumim time:         3.788
      Average time:         4.064

   Method: Sys5 Message
      Number of packets: 1000
      Maximum time:       163.764
      Minumim time:         2.966
      Average time:        35.344

---------------------------------------------------------
The test programs were almost identical: a single program
which fork()'d, the child wrote a bunch of packets, the
parent tried to read them.  The packets themselves contained
the current time from gettimeofday(2), which was compared to
the time after the read.

The message version used blocking read/write (msgrcv, msgsnd).
The UDP version used blocking write (sendto), and non-blocking 
read (recvfrom), initiated by a select().  To avoid queue
overflow, the UDP version slept 1 second between sends (the
select() was around to notify me if packets got lost - none
did).

The machine is an HP9000/360 (68030 @ 25MHz), running HP-UX
6.2.  A few users at the time, doing the sorts of things that
users do.

Anyone that wants the source code should mail me directly,
at 'keith%fstohp@crdgw1.ge.com'.

-kdg