FWIW, Chromium has run live user experiments testing performance with different per-host connection limits. Our current connection per host limit (6) reflects what we believe to be the most performant. On Sat, Mar 31, 2012 at 7:01 PM, Willy Tarreau <w@1wt.eu> wrote: > On Sat, Mar 31, 2012 at 12:44:45PM -0400, Peter Lepeska wrote: > > Watching this video, Patrick McManus makes a great point about increasing > > connections increases the likelihood that SYNs and SYN-ACKs will get > > dropped. That's a very convincing argument for me AGAINST my suggestion > to > > simply increase the per host limit. > > > > Thanks for sending the link, > > I agree with him on this too, and for having run tests with very large > per-host limits one year ago, I can say that these losses *do* happen. > What's worse is that they generally don't happen below a certain limit, > and that they happen by batches when you go beyond that limit. Of course, > this limit depends on the network between the two TCP endpoints and its > load and there is no real way to guess it. > > Willy > > >
Received on Sunday, 1 April 2012 21:12:22 UTC