I currently work with Prof. Dina Katabi on networking stuff. Before this I worked at Akamai as the senior architect for their load balancing and mapping system.
When packets are lost in 802.11 networks, often most of the bits are recieved fully correctly. This paper explores ways to take advantage of these partial packets to improve throughput. We find that in indoor stationary high SNR networks where auto-rate works fairly well, it's difficult it get more than about 20-30% performance improvement. While in outdoor or mobile situations where auto-rate struggles, taking advantage of partial packets can often improve performance by 2-3x.
In the push for higher wireless throughput for applications like HDTV, it seems inevitable that we'll be moving to use wider band radios which use more than the 20Mhz of current 802.11.
The problem with moving to wider bands is that the radio then either must use really low power, like UWB, which limits range or risk clobbering all of the other narrower band devices operating in the same region. SWIFT operates at 802.11-type power levels without clobbering other devices by doing two things:
First, it is able to intelligently learn almost the exact frequency range that a narrowband device is using by taking advantage of the fact that most narrowband devices will somehow react if SWIFT uses any part of their frequency band.
Secondly, through an intelligent PHY layer, it is able to take advantage of the many non-contiguous chucks of freqency space available between the narrowband interferers.
Resilient BGP Routing
Self-Consisent BGP: A Domain Dividied Against Itself Cannot Stand
Nate Kushman, Dina Katabi, and Bruce M. Maggs
Under Submission
This paper shows that more than half of the packet-loss caused by an inter-domain link failure can be traced back to iBGP. It presents SC-BGP, a modified version of iBGP which ensures that as long as at least one router in an AS continuously has a valid path to the destination, then no router in the AS will see BGP caused packet-loss. It also shows how SC-BGP can be combined with R-BGP to ensure continuous connectivity between all routers.
This paper presents R-BGP a modified version of eBGP which ensures continuous connectivity between ASes (although not within a single AS). I'm making the simulation code for this paper available on a limited basis, so e-mail me if you're interested in it.
This is a VoIP measurement study we did showing that about the half the periods of poor VoIP quality are highly correlated with BGP events, and could probably be avoided by a more resilent inter-domain routing protocol. This result helps motivate the solutions presented in the two papers above.
This was an early version of R-BGP (with the unfortunate name of C-BGP) which focused on strict consistency. After realizing that the overhead and delay caused by this was too burdensome, we decided instead on the lower overhead protocol of R-BGP.
In a previous life (before Akamai) I was a complier guy, and for my
Master's Thesis I showed that just by adding NOPS you can improve the
performance of the SPEC benchmarks by up to 9%. This was in reaction
to the common practice (at the time) of publishing compiler papers
presenting complicated complier techniques to achieve performance
improvements of only 5-10%.
Note: This web page layout was shamelessly stolen from Srikanth Kandula.
