On Tue, Jun 21, 2011 at 10:18:24AM -0700, Justin Erenkrantz wrote: > On Tue, Jun 21, 2011 at 9:57 AM, Greg Stein <gst...@gmail.com> wrote: > > That definitely implies something wrong on the client. If the server > > is slow, then the client should mostly be blocking. > > > > I'll investigate after I finish with this XML parsing thing in > > ra_serf. (which may fix the observed memory problem) > > As a quick data point: I ran with trunk of both SVN and serf - it > looks like sf.net turned off keep-alives over SSL - so every request > needs a new TCP connection. With trunk of serf (which has a bunch of > SSL and memory fixes), it takes ~15 minutes to do the checkout with no > memory growth that I could see...but, the constant re-opening of TCP > connections is pretty painful - but, it should be something that the > sf.net admins can resolve. -- justin
It's a bit worrying that our new default will require admins to tweak the server config to provide adequate performance for clients in the default configuration. Then again, there's the chicken-and-egg problem. If we don't force people to push for server-side config changes, they're never gonna happen... Though maybe we should wait for one more release cycle before making serf the default, and tell people that to get most out of httpv2 for 1.7 they should make sure their clients use serf and their server config is adjusted accordingly? That would provide some time for a soft transition. It might cause less pain for those who run into the problem with servers they do not control.
