[tbnl-devel] Firefox

[Edi Weitz]

On Fri, 30 Sep 2005 11:08:09 -0700, "Anthony F" <eatme444 at hotmail.com> wrote:

> Perhaps "Leak" was too strong a word.  I'm not sure exactly what is
> going on yet, but here are the symptoms:
>
> I start Apache, then TBNL.  I hit the test page with Firefox.  While
> watching the Process tab in Task Manager I follow a link.  Memory
> usage on the LispWorks process jumps anywhere from 2 to 5 megs.
> Following any other links within the test app or refreshing the page
> causes the same until I hit LispWorks Personal's heap size limit and
> it dies.
>
> I have yet to check if restarting Apache fixes the problem, but I
> will report on that later.  HOWEVER, I notice that even with IE a
> very small amount of memory seems to be leaking?  If you hit the
> Test app from IE, follow a link and/or repeatedly refresh I notice
> memory usage on the LispWorks process creep upwards ever so slowly.
>
> Having said this, I have not actually done any real investigating
> yet (like I said, I'm a total newb)... so this may or may not be a
> result of expected behavior.

I've just experimented with this myself a little bit and I have to
admit that TBNL on LispWorks does indeed seem to "leak" memory in that
the size of the LW application grows continuously without shrinking
again - on Linux as well as on Windows.  It turns out that a lot of
this is due to the external format code I introduced in 0.8.0 which
uses the FLI.  In particular, there's a call to SYS:IN-STATIC-AREA in
there which I guess is the culprit.  It was my understanding that
these statically allocated objects are subject to normal garbage
collection but it looks as if new static segments are allocated all
the time.  As the LW reference manual says: "Such additional segments
would cause the application to grow irreversibly."

But even pre-0.8.0 TBNL on LW "leaks" a bit.  It looks like at least
under heavy load the higher generations tend to grow so one should
probably call MARK-AND-SWEEP from time to time.

I'm a bit too busy right now to investigate this further but maybe
someone else wants to look at it and fix it.

> On a seperate note: I don't really want to start a mod_lisp
> vs. other options war, but is anyone using a Lisp server
> (AllegroServer, TBNL's server, etc) back end in a production
> environment?  I was wondering about performance, security, other
> issues, etc.  Personally, I feel "safer" using Apache as it's battle
> hardened and fairly ubiquitous, but maybe those feelings are
> unfounded?  Any comments?

I don't think there's a reason to assume that Lisp servers are
inherently less secure than Apache, actually I tend to think that
there are a lot of "typical" security holes that you'll see in a C/C++
app but not in a Lisp app.

As fas as performance is concerned I'm pretty sure that Lisp servers
are definitely "fast enough."  If one of the websites I'm maintaining
should really ever have so much traffic that it hits the performance
limits one can see when benchmarking Apache/mod_lisp/TBNL I'd be very
happy... :)

Having said that - there aren't many Lisp web servers out there that
fully implement HTTP/1.1 and you certainly want that if you're serving
production web sites.  TBNL stand-alone doesn't do HTTP/1.1 and
Araneida doesn't do it either AFAIK.  AllegroServer does, though, and
it has been used successfully in production environments like these:

  <http://www.pandorabots.com/>
  <http://www.franz.com/>

Another example for a Lisp web server used in a production
environment, although a bit dated, is the White House Electronic
Publications server which was running atop CL-HTTP when Clinton was
still in office.  CL-HTTP is also HTTP/1.1-compliant, BTW.  It just
has a somewhat peculiar license.

Cheers,
Edi.