Re: [GLIF tec] Buffering discussion

[Victor Reijs <victor.reijs@xxxxxxxxx>]

Hello Steve,

Sorry I was not at the GLIF meeting (personal issues), but I am
wondering what the outcome or progress of this buffering discussion was.

In some way I don't like to use bufferring, makes me remind me of the
ATm times;-). By the way the buffering needed is quiet small or not,
only some 10 MByte for a long 10 GE link, so in some way it is doable, I
assume this 10 MByte is independant of the number of active tcp flows
(or is it per ftp flow?).
{Sorry, perphas this all has been discussed at the meeting....)

Here at HEAnet we had some brainstorm stimulated by this buffering
e-mail, and we came up with a wild idea;-).
I will send this wild idea soon to this list (I already send it to
another list and some feedback has been goitten, so I will adjust the
text somewhat [wild ideas always need the corret wording;-)]).

Thanks for your summary.

All the best,


Victor




Erik-Jan Bos wrote:

> Tech Members,
>
> Upon request of Steve I hereby forward his message on the buffering 
> discussion.
>
> __
>
> Erik-Jan.
>
>
> > From: Steven Wallace <ssw@xxxxxxxxxxx>
> > Date: February 13, 2005 4:23:18 PM EST
> > To: tec@xxxxxxx
> > Subject: Fwd: Buffering discussion?
> >
> >
> >
> > Here's the gist:
> >
> > Thanks to Guy Almes for finding a small issue related to the correct
> > value of C for fiber and a big issue where I was converting to bytes
> > twice.
> >
> > What this graph is supposed to show is the buffering requirements of
> > a
> > 9,953,280,000 b/sec flow going over a mix of WAN-PHY and LAN-PHY
> > connectivity.  The speed difference between these two circuit types
> > is
> > over 300Mb/s, so buffering the difference in speed for the duration
> > of
> > a windows worth of data can be substantial.
> >
> > An Example:
> >
> > [Host A]----10GE LAN-PHY-------[15454]-----10GE WAN-PHY------[Host B]
> >
> >
> > So the fastest TCP can operate is limited to the slowest part of the
> > path: the 10GE WAN-PHY section (i.e. 9,953,280,000 b/sec).  The
> > window
> > size to support a WAN-PHY speed TCP transfer at 1000 miles is
> > 10,472,727 bytes.  Assuming the data is moving from host A to B, then
> > host A will be sending 10,472,727 bytes  (a window's worth of data)
> > at
> > line-rate toward host B.  The speed discontinuity between the LAN-PHY
> > and WAN-PHY sections of the circuit is 359,220,000 b/s, so for the
> > duration of the window transfer the 15454s must buffer this
> > difference
> > in speed.
> >
> > The following graph shows the buffering requirements for various
> > circuit distances and rate changes: