Known BEAM memory leak ? Profiling tools ?

Jon Holdsworth <>
Fri Jul 14 06:30:47 CEST 2000


Thank you Scott, and this could well be our culprit.

Jon


Scott Lystig Fritchie wrote:

> >>>>> "jh" == Jon Holdsworth <> writes:
>
> jh> To the erlang-questions people: Erlang running under Linux Redhat
> jh> 6.1 on Intel and Solaris 2.7 on Sparc, the BEAM program has been
> jh> seen to apparently leak memory after being left running for some
> jh> time.  It is possible that this 'memory leak' is in our own code,
> jh> but it has not revealed itself yet.
>
> Your own code?  Have you written any linked-in drivers (static or
> dynamic)?  Or is your code purely Erlang?  If the former, well, that's
> another bucket of mud.  The latter...
>
> ... are you sure you don't have a "process leak"?  By that I mean a
> number of processes which have escaped a supervision tree?  I ran into
> an evil memory consumption problem that, on the surface, looked like
> you describe: both "ps" and "top" said that the VM's UNIX process was
> eating a ton of memory, and I couldn't figure out why.  I suspected a
> memory leak with binary objects.
>
> It finally hit me to fire up the process manager, via "pman:start()".
> The silly thing timed out, causing virtual memory thrashing & wild
> disk activity.  After a few tries, I got it to pop up a window before
> the start function timed out.  Its process listing was *huge*.
> Dumbstruck, I tried "length(processes())", expecting about 150 and
> instead got I-forget-how-many thousand processes.  They weren't doing
> anything other than sitting around consuming a few kilobytes of memory
> each ... and blocking waiting to receive messages that would never
> arrive.
>
> Sure enough, we had a process leak.  Being much more careful about
> process linkage made the problem vanish.  ... Many thanks to Lennart
> Öhman for (much later) giving us a better and much more paranoid
> perspective on process linking: "Pretend that your system will run for
> 10 years straight.  Even a tiny process leak will cause a Siberia
> Problem(*) in that amount of time."
>
> Our worries about binary objects (in R6B-0) were unfounded.  I
> honestly don't recall why I didn't think about that kind of leak
> earlier, but I do remember how surprised I was when I found it.  And
> how chagrined I was to not think of it earlier.
>
> -Scott
>
> (*) When you need to fly/drive/hitchhike/walk all the way to an
> equipment closet in Siberia just to press a stupid reset button.




More information about the erlang-questions mailing list