From: Shawn Ostermann (sdo@picard.cs.ohiou.edu)
Date: 11/05/04
Subject: Re: tcptrace-bugs Re: AVL bug-fix available Date: Fri, 05 Nov 2004 09:00:21 -0500 From: Shawn Ostermann <sdo@picard.cs.ohiou.edu> Message-Id: <20041105140021.E14AE2EE2BF@picard.cs.ohiou.edu>
> > We have a bug-fix release of tcptrace, version 6.6.7, available from
> > the tcptrace web-site downloads page. The bug was in the function used
> > by the AVL Tree traversal algorithm to compare entries in the AVL Tree
> > with an incoming connection snapshot, in the event
> > of a hash collision.
> >
> > If you could let this release crunch on your huge
> > dumpfiles and let us know how it did, that would be very nice.
>
> No more infinite loops here, thanks!
Thanks for getting back to us. I'm still pretty skeptical that there
was actually a loop; the bug that we found wouldn't have caused that. I
suspect that it was a thrashing (constant paging) problem. I've seen it
hundreds of times in the past. We've organized that memory structure in
tcptrace to be as kind to VM as possible, but when a dataset gets
sufficiently large that it fills physical memory, performance falls off
quickly.
What I suspect was happening is that the bug that we found was making a
lot more connection records than necessary and causing your machine to
page so quickly that it didn't seem to be making progress.
After your note, we've spent a lot of time with HUGE trace files the
last few days. The one we were playing with yesterday was only 800MB,
but it was great for chewing up memory. Once we put it on a machine
with enough physical memory, it ran in about 50 seconds.
Shawn
This archive was generated by hypermail 2.1.7 : 11/05/04 EST