[plt-scheme] Massive memory consumption

Greg's right, though a caveat and explanation. I ran the memory tests with
no timeouts on the web server's continuations. This was done because the test
was on Continue, which is designed to run for only a few months at a time
and a goal is to have no timeouts for its entire run.

So this was a test
against cleaning up bits of temporary memory, but memory use is expected to
go up over time. 3m may be a better choice for long-running web programs in
general, since it should clean up all unreachable memory, but for my purposes
Boehm seems to be better.

The 3m collector had slightly higher memory use
and significantly lower performance when memory grew tight. After 1500 hits:


Boehm: 272Mb
3m: 309Mb

Average response times for hits 1200-1500:
Boehm: 4.68 secs
3m: 28.95 secs


This was on a Mac OS X machine with
512Mb of RAM. The performace hit for 3m was, I believe, directly related to
thrashing of the OS's virtual memory system. Not every 3m hit took 28 seconds,
but page-in delays made some hits take such a long time that it brought the
average up. I ran these tests last Spring with (I think) the latest official
source release (v207, I believe).


-- Pete


--- Greg Pettyjohn <gregp at ccs.neu.edu
wrote:
  For list-related administrative tasks:
>   http://list.cs.brown.edu/mailman/listinfo/plt-scheme

> 
> Prior to 299, Pete Hopkins ran some stress and, if I recall,
> the
short answer seemed to be that the 3m collector did not
> work as well with
the web-server as did the Boehm collector.
> On Oct 30, 2004, at 12:12 PM,
David J. Neu wrote:
> 
> > This should have gone to the list:
> >
> >
Matthew,
> >
> > Would using 3m with the PLT web-server be a good strategy?

> >
> > Thanks,
> > David
> 
>