[Search for users] [Overall Top Noters] [List of all Conferences] [Download this site]

Conference noted::atm

Title:atm
Moderator:NPSS::WATERS
Created:Mon Oct 05 1992
Last Modified:Thu Jun 05 1997
Last Successful Update:Fri Jun 06 1997
Number of topics:970
Total number of notes:3630

894.0. "cell transfer time cost?" by NETRIX::"eric gan@bejvc" (eric gan) Tue Apr 08 1997 06:21

1) How many time cost does one cell transfer from one ATM port to another?
2) Which is the best solution for parallel processing? We have two Alpha 4100
servers
need to working together on network, which network technology is enough to
avoid bottleneck? ATM?
FDDI? F.E.? or maybe switch ethernet is enough!? :-)

any comments will be wellcom,
thanks,
eric

[Posted by WWW Notes gateway]
T.RTitleUserPersonal
Name		DateLines
894.1ATM and FDDI *switch* latency is the sameNPSS::WATERSI need an egg-laying woolmilkpig.Tue Apr 08 1997 09:2214
  The latency through GIGAswitch/ATM is 12 microseconds (first bit in to
  first bit out).  The latency through GIGAswitch/FDDI is roughly the same
  --usually quoted as 15 microseconds.

  I don't know by how much the latency of ATM and FDDI host interfaces
  differ.  Anyway, unless you're using special protocols for parallel
  processing (such as PVM over raw ATM), the latency will be dominated by
  the host driver and protocol stack (context switching and all that).
  Message latency from one application to another, through TCP and either
  FDDI or ATM, can be 300 microseconds or more.

  That DIGITAL's ATM LANs suffer no congestion losses may help with some
  workloads.  There is a high-performance-computing Notes conference where
  you may get a better answer.  --gw
894.2Re: cell transfer time cost?QUABBI::"[email protected]"Hal MurrayWed Apr 09 1997 21:3946
There are a bunch of graphs on my web pages.  Start at:
  http://src-www.pa.dec.com/~murray/an2-perf/perf.html

The 750/350 takes about 28 uSec to send a 1 cell packet to itself.
That's the same as the time it takes to send a packet from one
machine to another.  Multiply by two for round trip times.
  http://src-www.pa.dec.com/~murray/an2-perf/host-latency.html
The fine print will depend upon how busy the memory/IO system is.

That's the best you will get.  I worked real hard to measure just
the raw hardware times.  No interrupts or anything like that.

Similar graphs for sending packets through switches are in:
  http://src-www.pa.dec.com/~murray/an2-perf/switch-abr.html


For small packets, the delay is primarily software.
For WAN links, most of the delay will be speed of light.

There are TCP and UDP request/response times in:
  http://src-www.pa.dec.com/~murray/an2-perf/tcpudp-batch.html
That's the round trip time, user-user-user for a well tuned
program that doesn't do anything sleazy.  Divide by two to
get the one way timings.  Kernel-kernel transfers would be faster.

Bottom line is roughly 1/4 to 1/3 of a ms for a modern machine
sending tiny packets.


----

I have some rough FDDI data - it isn't linked into the main page.
  http://src-www.pa.dec.com/~murray/an2-perf/tcpudp-batch-fddi.html

Request-response looks slightly faster for small packets, but
that's just a rough estimate.

----

For large transfers, you can get close to 100 megabits on switched
FDDI and close to 135 on ATM over OC-3.

If your FDDI isn't switched then you get to round down if other
machines are using the net.  (You don't actually need a switch
to get "switched" behaviour if you only have 2 machines.)
[posted by Notes-News gateway]
894.3thank you very much, BTW ...NNTPD::"eric gan@bejvc"eric ganTue Apr 22 1997 01:0111
Murray,

Thank you very much, your information help me a lot!

BTW, are you useing the "news" application to post notes? So, could you
tell me how to config the news' pereferance, and which gateway or proxy
server I should use?

Thanks a lot,
eric
[Posted by WWW Notes gateway]