Vijay,
On Wed, 31 Mar 1999, Vijay G Bharadwaj wrote:
> On Wed, 31 Mar 1999, Eric Travis wrote:
>
> > Only if the definition of 'decent' translates to an effective
> > cwnd of ~1 or 2 segments :o(
> >
> > You really want your (non-congestion based) losses bunched within
> > single RTTs that are fairly widely spaced [how wide will depend
> > on what your optimal cwnd for the path would be]. Otherwise TCP
> > will be severely under driving your link.
>
> Do you have data to support this? I have lots of measurements over a
> channel simulator that say I can get >90% link utilization on a T1 using
> TCP+standard_gadgets at 1E-8 BER, around 85% at 1E-7, and just over 50% at
> 1E-6. So even at 1E-6 I'm getting way better than your 1-2 segment cwnd
> would give me. These are random errors, delay is GEO (250ms). If you have
> data to the contrary I'd be very interested.
I do have some tcpdumps of pathological runs tucked away - I'll dig for
them. However, I'm not so sure your data contradicts me... My errors
were intentionally whacking a segment every nth frame (set up the
channel emulator to do this for me). For T1 rates, your numbers don't
surprise me - they don't contrict me either :o)
BER is the wrong metric to be using here. What is important to TCP is
segment loss distribution.
While you have introduced random errors at up to 1E-6, can you translate
that into a frame loss rate/distribution? What is the distribution of the
errors introduced? I'm confident that you are *not* losing one (or more)
segments every one or two RTTs. Have you correlated your loss events to
specific RTTs?
To get any kind of decent link utilization you *must* be growing
your congestion window *far* more often than you are halving it.
Again, I'll dig for my traces and post them when I find them.
> > I'm confident that if one pulls the archive and does a
> > grep on BER, errors or loss - you will be able to extract
> > virtually all the collected wisdom/opinions from the cyclic
> > (and often passionate) "BER Wars" thread that appear every
> > 2-3 months during 1998. I'm pretty sure it is all there
> > already (several times).
>
> I just did this. A search on BER gave maybe 6-7 hits, mostly incidental
> stuff, and two of them were actually because someone had ROBERT in their
> email address. Maybe this list needs a FAQ? ;) I'll look harder, but just
> FYI...
Bummer - let me pull the archive and search for errors and loss to check;
Those discussions occured like clockwork. :o)
A FAQ would be an excellent thing though!
Eric
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:54 EST