wiki:TahoeTwo

Version 4 (modified by davidsarah, at 2009-10-19T18:25:09Z) (diff)

--

Copied from http://tahoebs1.allmydata.com:8123/file/URI%3ACHK%3Ayhw67uwzcdszqtzim6zxmiqzte%3A323pu5fz27sbntnxgffa57eqisabggvdnmq3zqoqyckgkimgmfma%3A3%3A10%3A3448/@@named=/peer-selection-tahoe2.txt

THIS PAGE DESCRIBES HISTORICAL DESIGN CHOICES. SEE docs/architecture.txt FOR CURRENT DOCUMENTATION

When a file is uploaded, the encoded shares are sent to other peers. But to which ones? The PeerSelection algorithm is used to make this choice.

Early in 2007, we were planning to use the following "Tahoe Two" algorithm. By the time we released 0.2.0, we switched to "TahoeThree", but when we released v0.6, we switched back (ticket #132).

As in TahoeThree, the verifierid (= StorageIndex) is used to consistently-permute the set of all peers (by sorting the peers by HASH(verifierid+peerid)). Each file gets a different permutation, which (on average) will evenly distribute shares among the grid and avoid hotspots.

With our basket of (usually 10) shares to distribute in hand, we start at the beginning of the list and ask each peer in turn if they are willing to hold on to one of our shares (the "lease request"). If they say yes, we remove that share from the basket and remember who agreed to host it. Then we go to the next peer in the list and ask them the same question about another share. If a peer says no, we remove them from the list. If a peer says that they already have one or more shares for this file, we remove those shares from the basket. If we reach the end of the list, we start again at the beginning. If we run out of peers before we run out of shares, we fail unless we've managed to place at least some number of the shares: the likely threshold is to attempt to place 10 shares (out of which we'll need 3 to recover the file), and be content if we can find homes for at least 7 of them.

In small networks, this approach will loop around several times and place several shares with each node (e.g. in a 5-host network with plenty of space, each node will get 2 shares). In large networks with plenty of space, the shares will be placed with the first 10 peers in the permuted list. In large networks that are somewhat full, we'll need to traverse more of the list before we find homes for the shares. The average number of peers that we'll need to talk to is vaguely equal to 10 / (1-utilization), with a bunch of other terms that relate to the distribution of free space on the peers and the size of the shares being offered. Small files with small shares will fit anywhere, large files with large shares will only fit on certain peers, so the mesh may have free space but no holes large enough for a very large file, which might indicate that we should try again with a larger number of (smaller) shares.

When it comes time to download, we compute a similar list of permuted peerids, and start asking for shares beginning with the start of the list. Each peer gives us a list of the shareids that they are holding. Eventually (depending upon how much churn the peerlist has experienced), we'll find holders for at least 3 shares, or we'll run out of peers. If the mesh is very large and we want to fail faster, we can establish an upper bound on how many peers we should talk to (perhaps by recording the permuted peerid of the last node to which we sent a share, or a count of the total number of peers we talked to during upload).

I suspect that this approach handles churn more efficiently than TahoeThree, but I haven't gotten my head around the math that could be used to show it. On the other hand, it takes a lot more round trips to find homes in small meshes (one per share, whereas TahoeThree can do just one per node).