Re: [Monotone-devel] BSD Buildbots

[William Uther]

On 03/02/2008, at 9:55 PM, Lapo Luchini wrote:

> Zack Weinberg wrote:
> > I think there might be something else weird with those buildbots.
> > That test shouldn't use anything like 256MB of memory - in fact,
> > massif claims it never gets above 10MB (see attached graph).  It does
> > repeatedly allocate and then free about 10MB of memory, though; the
> > only thing I can think of is that the malloc() implementation on the
> > buildbot is stuck in some debugging mode where it never reuses a
> > memory block, and that it has substantially more overhead than
> > massif's (fairly dumb) guess.

The graph in the email you gave showed that it was new and delete  
allocating
the most memory, not malloc and free.  That might be another place to  
look
for debugging options.  Are any special command like options used when
compiling?

> Indeed "time -l" confirms that (and AFAIK that reads kernel data about
> actual max memory used, it's not an approximation).
>
> % /usr/bin/time -l softlimit -d11000000 sh -c './unit_tester
> refiner:various_counts 2>/dev/null'
> load: 0.99  cmd: unit_tester 49383 [runnable] 312.86u 3.36s 98% 9704k
>      602,70 real       573,98 user         6,34 sys
>     10768  maximum resident set size
>      4564  average shared memory size
>      6080  average unshared data size
>       131  average unshared stack size

Some questions...

Firstly, I note that 10768*1024 = 11026432.  That means that the  
maximum resident set size is greater than the softlimit you set, no?   
Do we know if the unit test succeeded in this test, or did it die in  
an OOM?  Can you try a simple:

/usr/bin/time -l './unit_tester refiner:various_counts 2>/dev/null'

without the softlimit?

Oh, and what does "ulimit -a" return?

Secondly, I note that the number time returns is labelled "maximum  
resident set size".  Is this including swap or not?  I found this web  
page an interesting read: http://www.perlmonks.org/?node_id=626693

Cheers,

Will        :-}