On 14/05/2016 12:38 a.m., Garri Djavadyan wrote: > On Fri, 2016-05-13 at 08:36 +1200, Amos Jeffries wrote: >> Have you given collapsed_forwarding a try? Its supposed to prevent >> all >> the duplicate requests making all those extra upstream connections >> unti >> at least the first one has finished getting the object. > > Amos, I believe that the above quote describes default Squid's action, > which does not require collapsed_forwarding. The details of my > experiments can be found here http://bugs.squid-cache.org/show_bug.cgi? > id=4511#c0. Thanks. > The default action should be to fetch each range request separately and in parallel. Not caching the results. When admin has set only the range offset & quick-abort to force full object retrieval the behaviour Heiler mentions happens - lots of upstream bandwidth used for N copies. The first one to complete starts to be used as a HIT for future reuqests. But as each of the initial transfers completes it replaces the previously cached object as the one being hit on. So timing is critical, if Squid happens to delay any of the parallel requests just long enough in its TCP accept() queue, auth or ACL processing they could become HITs on an already finished object. Amos _______________________________________________ squid-users mailing list squid-users@xxxxxxxxxxxxxxxxxxxxx http://lists.squid-cache.org/listinfo/squid-users
