On Mon, Feb 15, 2010 at 7:25 AM, Matthew Toseland <toad at amphibian.dyndns.org> wrote: > On Sunday 14 February 2010 04:29:31 Evan Daniel wrote: >> On Sat, Feb 13, 2010 at 7:00 PM, xor <xor at gmx.li> wrote: >> > First of all, I like your calculations very much and I wonder why nobody >> > calculated this before FEC was implemented. If I understood this correctly >> > then a 700mib file with block success rate p=0.58 will have a 48% total >> > success chance. This sucks... >> >> Thanks :) >> >> That's correct. ?However, in some ways it isn't *quite* that bad. >> There's a reason I picked p=0.58: it's nice and dramatic. ?But, the >> curve is fairly steep there :) ?At p=0.60, it's 92% success rate. ?At >> p=0.61, it's 97%. >> >> There are two ways to look at the success rate. ?One option is to pick >> some block success rate and look at the resulting file success rates >> for a variety of files. ?By this metric, the interleaved coding is a >> stunning improvement. ?The other approach is to pick a target file >> success rate, and see what block success rates are required to get it >> for different scenarios. ?So at 700M, we need p=0.61 to get 97% with >> simple segments. ?With interleaved coding, we need p=0.56. >> >> The latter approach is actually probably more directly meaningful. >> The improvement from interleaved coding is significant, and useful, >> but modest. ?What it says is that instead of being able to lose 39% of >> blocks and still expect to recover the file, we can now expect to >> recover the file after losing 44%. ?If we assume that individual >> blocks are well modeled as having a half-life to disappearance, then >> it takes 1.17 times longer for the file to become inaccessible if we >> use the interleaved coding. > > However, in practice, it appears that many files do stall at close to 100%. > This could be a client layer bug, but it could be that we have lost a segment.
If it's a lost segment, these changes might not change the perception much. If 20% fewer downloads fail, people might not really notice. And the observed behavior of stuck at 99% might not change much. Evan Daniel
