Bram Cohen Reacts To Microsoft Avalanche

June 21st, 2005 · No Comments

Bram Cohen

Bram Cohen (Bittorrent creator) has written a piece on his blog in which he publishes his thoughts to Microsofts supposed bittorrent replacement, Avalanche. As you can probably guess, he’s less than enthousiastic:

One thing badly missing from this paper is back-of-the-envelope calculations about all of the work necessary to implement error correction. Potential problems are on the wire overhead, CPU usage, memory usage, and disk access time. Particularly worrisome for their proposed scheme is disk access. If the size of the file being transferred is greater than the size of memory, their entire system could easily get bogged down doing disk seeks and reads, since it needs to do constant recombinations of the entire file to build the pieces to be sent over the wire. The lack of any concrete numbers at all shows the typical academic hand-wavy ‘our asymptotic is good, we don’t need to worry about reality’ approach. Good asymptotics are one thing, but constant multipliers can be killer, and it’s necessary to work out constant multipliers for all pontentially problematic constants, not just the easy ones like CPU.

The really big unfixable problem with error correction is that peers can’t verify data with a secure hash before they pass it on to other peers. As a result, it’s quite straightforward for a malicious peer to poison an entire swarm just by uploading a little bit of data. The Avalanche paper conveniently doesn’t mention that problem.

As you’ve probably figured out by now, I think that paper is complete garbage. Unfortunately it’s actually one of the better academic papers on BitTorrent, because it makes some attempt, however feeble, to do an apples to apples comparison. I’d comment on academic papers more, but generally they’re so bad that evaluating them does little more than go over epistemological problems with their methodology, and is honestly a waste of time.

If you’re interested in doing more fleshed out research on error correction in BitTorrent, I suggest starting with a much less heavyweight approach. Having peers transfer the xor of exactly two pieces could potentially get most of the benefits of full-blown network coding.

Hehe….this one’s straight from the horses mouth