SCTP Discussions

Text archives Help


Re: SV: [SCTP-Discussion] SCTP performance measure


Chronological Thread 
  • From: Thomas Dreibholz <dreibh AT exp-math.uni-essen.de>
  • To: Torbjörn Andersson <torbjorn.andersson AT kau.se>
  • Cc: "'Michael Tuexen'" <Michael.Tuexen AT micmac.franken.de>, <discussion AT sctp.de>
  • Subject: Re: SV: [SCTP-Discussion] SCTP performance measure
  • Date: Mon Jul 26 10:27:02 2004
  • List-id: SCTP Discussions <discussion.sctp.de>

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Am Montag, 26. Juli 2004 09:54 schrieb Torbjörn Andersson:
> Hello everyone. You are having an interesting conversation here and I
> have some questions.
>
> See statements and/or questions below.
>
> > -----Ursprungligt meddelande-----
> > Från:
> > discussion-admin AT sctp.de
> >
> > [mailto:discussion-admin AT sctp.de]
> > För
> > Michael Tuexen
> > Skickat: den 23 juli 2004 14:45
> > Till: Andreas Jungmaier
> > Kopia: Le Vinh Dat;
> > discussion AT sctp.de
> > Ämne: Re: [SCTP-Discussion] SCTP performance measure
> >
> > Just to be clear:
> >
> > What I meant was that the combination of the sctplib
> > with the socketapilib might not be very good for
> > performance measurements. You can do it with the
> > sctplib only.
> >
> > The reason is that the socketapilib use threads...
>
> However, is the timing of your implementation accurate? In terms of the
> accuracy of the retransmissions and more.

Yes, it is accurate.


> Naturally, the sctplib must have the necessary computing resources
> available but today computers have no problems with that, or what do you
> say?

Yes, a fast computer will have no problems saturating a 100 MBit/s or even a
Gigabit link.

But when you compare the amount of CPU cycles necessary to transmit this
amount of data using kernel TCP and our userland implementation, you will of
course notice that kernel transport requires less CPU cycles.


> A protocol performance measurement is in my eyes not a test in an
> unlimited network (100mbit/s). The performance of a protocol is in my
> eyes seen during congestion and other losses.
>
> So if the retransmission accuracy of your implementation is ok then is
> the implementation valid for performance tests, at least in networks
> with limited bandwidth and buffering resources.
>
> Do you agree with me?

Yes, if these points are the goal of your performance tests, a userland
implementation will be fine.


Best regards
- --
=======================================================================
 Dipl.-Inform. Thomas Dreibholz

 University of Duisburg-Essen                  Room ES210
 Inst. for Experimental Mathematics              Ellernstraße 29
 Computer Networking Technology Group            D-45326 Essen/Germany
- -----------------------------------------------------------------------
 E-Mail:    
dreibh AT exp-math.uni-essen.de
 Homepage:   http://www.exp-math.uni-essen.de/~dreibh
=======================================================================
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.4 (GNU/Linux)

iD8DBQFBBMA132BbsHYPLWURAhNMAKCdOXTX4QjDdBfdKG2OZ+WyF0mxvwCgmWzF
feHdqrHKuLZkZ3WQOWsnzWo=
=IVK5
-----END PGP SIGNATURE-----




Archive powered by MHonArc 2.6.19+.

Top of page