Path: utzoo!attcan!utgpu!jarvis.csri.toronto.edu!cs.utexas.edu!samsung!rex!ames!sgi!calcite!vjs
From: vjs@calcite.UUCP (Vernon Schryver)
Newsgroups: comp.unix.i386
Subject: Re: Shrink Wrapping X Windows Clients on Intel Hardware
Message-ID: <80@calcite.UUCP>
Date: 17 Dec 89 03:59:14 GMT
References: <1989Dec8.005923.23810@pcrat.uucp> <129232@sun.Eng.Sun.COM>
Organization: Rhyolite Software, Mountain View, CA
Lines: 37

In article <129232@sun.Eng.Sun.COM>,plocher@sally.Sun.COM (John Plocher) writes:
> ...
> One workaround may be to set 
> 	DISPLAY=localhost:0
> and force the client to use the loopback network interface
> for connections to the server.  Slow, but faster than not
> working at all :-)

How slow is that?  That is, how does the loopback interface perform
on common 386 UNIX'es, measured either by something in X said to be nasty
like ico, or a pure TCP/UDP measure like TTCP, but <<not>> FTP, since I'm
asking about the network and not the file system.

Fast workstations (e.g. based on MIPS or SPARC) get 1.5-2.0 MByte/sec thru
TTCP via loopback.   I've been told WIN TCP on a 386 clone gets a
respectible >300KByte/sec thru TTCP on an ethernet.  How do it and the
others do via loopback?  Since no real I/O is involved, this is a measure
of how much overhead is in the TCP implementation, as well as how fast the
CPU can bcopy or otherwise shuffle bytes, as well as other stuff like
the scheduler.

I would be happy to mail anyone who can't FTP from BRL or the other places
a copy of ttcp.c.  It is supposed to appear in comp.sources.unix someday.

> ...[distressing but no doubt accurate description of the state of the
>	i386 network world deleted]...
> This is a win when you realize that the library that the client
> actually uses is the one provided with the server that is resolved
> at run time!

How much will/does Sun/SVR4-style dynamic linking cost in performance?
I've heard rumors ranging from 3% to 15%.  What are real numbers, in
a variety of situations?


Vernon Schryver
vjs@calcite.uucp