Bandwidth breakthrough from MIT

[gturner]

Up to a ten fold improvement in data rates over lossy networks using the same bandwidth, just by using linear algebra to encode a series of packets as a formula that allows the reconstruction of dropped packets.

http://www.technologyreview.com/news/429722/a-bandwidth-breakthrough/

Unfortunately the article is vague on the details, but it sounds promising.

 

[Robert Maxwell]

I read about this recently. Really cool idea--makes a lot of sense for high-loss environments like wireless networks. What I think is novel here is that most data integrity checking has historically been limited to determining whether the data is correct, not rebuilding it once it is lost (except from the original source.) Including some extra data to permit reconstructing missing pieces is pretty clever and novel.

 

[Lindley]

The concept of using linear combinations of packets to reconstruct missing ones has been around for a few years now. What's new here is that they've found a way for intermediate routers (as opposed to the original source node) to re-encode the packets in such a way even when they don't have all of them.

 

[Robert Maxwell]

The concept of using linear combinations of packets to reconstruct missing ones has been around for a few years now. What's new here is that they've found a way for intermediate routers (as opposed to the original source node) to re-encode the packets in such a way even when they don't have all of them.

That makes sense. And I'm sure they are keeping the exact details secret so it can't just immediately be open-sourced and given away for free to all hardware makers.

 

[gturner]

Yep, but I'd love to know at least the basic approach they're taking.

Error detection and correction has been around a long time, but usually focused on a flipped bit, or a couple of them. We even used to deal with it at the code level back when actual BBS's were around and we'd transfer files with things like Z-modem or X-modem.

So are they doing something like a logical extrapolation of the earlier techniques to cover entire packets, or something else entirely? Does the method depend on, say, scattering enough data across all the packets to cover a 5% loss rate with 5% bigger individual packets, or what?

Needless to say, their solution has probably already been hacked and somebody is probably incorporating it into the next generation of bitTorrent clients.

 

[Lindley]

From what I read in the comments section (there are several extensive posts by the original investigators), it sounds like their technique is best suited for installation in router network stacks rather than endpoint applications.

 

[Geckothan]

Forward error correction is not new :/