jbenet
commented
9 years ago I'll try to read this in next couple of days. but skimming it, this isn't a good paper. they collected very little data and tried very few things. compare to, for example, papers like these
- http://www.cs.helsinki.fi/u/jakangas/MLDHT/
- http://www.cs.umd.edu/~dml/papers/bittorrent_sigcomm08.pdf
- http://web.mit.edu/keithw/www/Learnability-SIGCOMM2014.pdf
i wouldn't adopt any of their conclusions, but definitely interesting to see how they modeled the problem.
Firewalled hosts were often faster for some unknown reason
probably because they were nodes likely to be in corporate or university networks, with fast connections :)
http://iptps03.cs.berkeley.edu/final-papers/adaptive_selection.pdf
This paper seems to solve a similar problem as bitswap-ml, but in the context of the Gnutella network.
The features they used were:
Using these attributes they created a decision tree to rank peers on a discrete scale from "Very Likely Slow" to "Very Likely Fast". They also found that having the busy flag set was highly correlated with poor connections, so they gave busy peers their own rating, "B", which is a rank lower than Very Likely Slow, and did not include it in the decision tree. Some interesting results they found when creating the tree were:
Once potential peers were ranked using the decision tree, the researchers used a Markov decision process to intelligently perform partial downloads to find the best peers in minimal time.