File Swarming Systems

New: The paper on coupon replication systems won the ACM SIGMETRICS 2005 Best Paper Award

Overview

We refer to file swarming systems as a general paradigm of file distribution where file is sliced into chunks; user is granted initial chunk from a server; and other chunks are replicated among users.

 

Example: BitTorrent

 

A benefit of the swarming paradigm is better scalability compared to standard central server solutions. Swarming may also better utilize network links, thus achieve larger throughput.

 

System is characterized by replication strategy that consists of (i) peering strategy (who to exchange from ?)  and (ii) chunk selection strategy (which chunk to pull ?) For example, BitTorrent uses as replication strategy a mixture of local, greedy and random choices. Choice of replication strategy may significantly effect the file download time.

 

Performance of file swarming systems is believed to critically depend on user altruism in uploading chunks even after have downloaded all chunks of a file (acting as seeds). It is believed that if users were not altruistic, then some chunks may become rare and this would severely increase the time for a user to acquire all file chunks. This problem may be mitigated by some smart choice of replication strategy, for example, rarest chunk first.

 

We provide analysis of file swarming based on a model: we call coupon replication systems.

 

We consider simple-minded replication strategies. The peering strategy is either (layer) pick a random user from those having same number of coupons as the instigator (flat) pick a random user. Once a peering is made, the instigator pulls a random chunk of interest from the encounter user.

 

We assume users are non altruistic.

 

The results suggest: already simple random replication strategies yield good performance. The choice of either replication strategy or user altruism may not be critical.

 

References

Tutorials & lectures

  • New: File Swarming Systems, Lecture, OPT 012 Operations Research Lecture, Charles University, Prague, Czech Republic, Dec 2005.
    1. Slide show pps
    2. Last-chunk experiments zip (download and run lastchunk.m with Matlab)

People

Laurent Massoulie, Milan Vojnovic

 

Related work

Models

Measurements

 

May 04, 2006, http://research.microsoft.com/~milanv