swift status report
DESCRIPTION
SWiFT Status Report. Dan Berger 060203. Outline. Split TCP: The Aftermath False Link Failures: ns-2, The Revenge Current Work: 802.11 Rendezvous Conclusion. Split TCP: The Aftermath. Recall my last update: - PowerPoint PPT PresentationTRANSCRIPT
Outline
Split TCP: The AftermathFalse Link Failures: ns-2, The RevengeCurrent Work: 802.11 RendezvousConclusion
Split TCP: The Aftermath
Recall my last update: SplitTCP, while intuitively attractive; did
not perform as expected when modeled accurately.
Since then a handful of additional experiments were performed to strengthen understanding.This work is now indefinitely on hold.
False Link Failures
Subsequent discussions focused attention on false link failures. Incorrect link abandonment due to
congestion.
Extensive simulations were performed: Static string of various lengths > 3* Looking at the number and distribution
of false link failures
* This becomes important – stay tuned
ns-2, The Revenge
Just when you thought it was safe to run a simulation…Analysis of the simulation results revealed surprising patterns. FLF’s were occurring with almost
predictable periodicity. Theories abounded: Long range dependence? Invaders from Planet X?
A Mystery Solved
A simulation of a three node provided a key insight: FLF’s shouldn’t occur – but were.
Something was rotten in Denmark.After consultation with Vikram, we isolated the problem to a recent (2/03) change to the NS 802.11 code.
Nothing to see here, move along…
A fix was developed and tested – once we were satisfied it was submitted to the ns-2 maintainers.Unfortunately they still haven’t applied it to the official sources.It’s available on my cs web page.
Back to your regularly scheduled programming
Attention was returned to False Link Failures. They’re not, as far as we can tell,
periodic But they are hugely destabilizing.
Eliminating* FLF’s saw a 100% improvement in goodput.
*Via a static topology, simulation-only hack
Current Work: Rendezvous
During discussion, Prasun suggested modifying the MAC to exchange packets in an effort to reduce channel contention. Total data transmission time is bounded.
An initial implementation has been done, and showed positive results (30% improvement @ 20 hops).
Rendezvous802.11 802.11 Rendezvous*
RTS
CTS
Data
ACK
RTS
CTS
ACK
Data(1)
Data(2)
ACK
Rendezvous: Current State
Similar results were published in ACM Mobile Networks & Applications in 2000. We’re assessing the appropriate
reaction.
The model has been iterated on once, and has (at least) one more iteration before it’s general.
Conclusion
This, and similar work by Frank is being targeted for an Infocomm paper.
Questions?