Deep Diving into BitTorrent Locality
This paper looks at P2P traffic over bittorrent from a large database of torrents. The paper considers the effects of localising bittorrent traffic on performance and ISP cost saving.
Data: The data set is one of the impressive things about this paper. 100K torrents of which 40K active. Demographics from 3.9M concurrent users and 21M total users over a day from 11K ISPs. Speed test results from ookla and iplane.
Chunk dispersion strategies for Bittorrent are tried: Random, Locality and LOIF (Locality only if Faster) and “Strict” (only local peers allowed plus at most one remote peer). Note that because of the unchoking mechanism of bittorrent, Random already induces some measure of locality (higher capcity peers are selected preferentially and local peers are likely to be higher capacity). This is assessed in section 3. Two modes are considered.
1) In “sparse mode” all nodes outside local ones (ones within the same ISP) have very different speeds. A simple model is created for selecting enough local nodes from a random draw of nodes based on selecting at random and getting at least local.
2) In “dense mode” local and remote nodes have simmilar speeds. Random selection will not localise traffic. (No intermediate state seems to be considered).
Locality is a policy which when the node asks for nodes provides as many local ones as possible. Obviously this has more benefits in dense mode than sparse mode. Improvement factors (in proportion of local nodes selected) versus sparse and dense mode are given based on simple probabilistic assumptions.
The concept of “inherent localizability” is developed using a value which captures the speed difference which can be tolerated between two nodes which mutually unchoke. An assumption seems to be made here is that the speeds between two ISPs and are equal regardless of the nodes (that is, it is the inter domain costs which predominate rather than intra domain costs or local bottlenecks). This localizability seems very unpredictable (fig 5). Some torrents are considered “unlocalizable”.
A european and an US ISP are analysed in detail. The US ISP has predominantly English language content in torrents and this has global reach in interest. In the European ISP (non English speaking) the popularity distribution of torrents is very different from the global average and the US ISP (which has a similar popularity distribution to the global average).
A model is created for seeders and leechers in networks (apparently this ratio is similar across all networks studied). Experiments are carried out incorporating demographics from torrent analysis, speed distribution (either all nodes use median from country or iPlane speeds used) as last mile bottlenecks and seeder/leecher ratios. Transit traffic reduction versus random is compared as is reduction of median QoS.
Sidenote: figure 4 shows a CDF of speeds of upload from iPlane and from ookla. These are very different but I can't find an explanation of why (iplane seems to give much higher speeds).
Validation is performed by integrating LOIF, Locality and Strict into the mainline bittorrent client (v 5.2.2) and testing the number of neighbours local or otherwise.
The paper concludes that locality can yield win-win (reduced transit traffic and improved QoS for users) but “unlocalizable torrents” provide obstacles.
Page generated 2014-01-07, by jemdoc.