RE: [pooma-dev] Profiling POOMA: How to? (1/3)

[James Crotinger]

Gaby, Microsoft Outlook has no idea what your attachments were. Here is a snippet of what I received:

----------------------------------------------------------------------
Hmm, I'm embarrased because I don't (yet) have any idea of why you're
seeing that behaviour.  Which parameters did you run the ABC
test with?  I tried the following to see whether I'll get a core dump

        ./ABC --run-impls 2 --sim-params 1 3 100

and the test completed fine, but I strongly suspect I must be testing
with the wrong parameters (the "nan" part is a bit intriguing).  The
output is in ABC.out and the profiling information (flat profile and
call graph) is in ABC.prof, all attached.

I'm using GNU gprof-2.10.91 on a linux box.

Thanks,

-- Gaby

--Multipart_Fri_Aug_24_02:46:57_2001-1
Content-Type: application/octet-stream
Content-Disposition: attachment; filename="ABC.out"
Content-Transfer-Encoding: base64

Ci4vQUJDIEJlbmNobWFyawotLS0tLS0tLS0tLS0tLS0KUnVubmluZyBzYW1wbGUgIzEgZm9y
IFBvb21hSUkgQnJpY2sgSW1wbGVtZW50YXRpb246CiAgTiA9IDEuLi4KICAgIENvcnJlY3Ru
ZXNzIHRlc3QgdmFsdWUgZm9yIE4gPSAxIGlzIG5hbi4KICBOID0gMS4uLgogICAgQ29ycmVj
-------------------------------------------------------------------------

Regarding your questions, it wasn't ABC that crashed, it was gprof when I ran it on ABC's gmon.out file (or whatever it is called).

The nan's you're seeing for the above args are probably related to the input parameters indirectly - the "correctness value" that is printed out for this test is the value at the point "N/2,N/2". For reasons that escape me, the domain for the arrays in this test is [1..N]x[1..N]. So if N == 1, N/2 give 0 and this results in the correctness calculation trying to read a(0,0), which is out-of-bounds. (This would be caught with bounds checking on, if we still have such a thing 8-). The upshot is that N == 1 is a bad value. Also, anything less than N == 10 will not work with the default number of patches for the UMP tests since they try to decompose into npatch x npatch patches, which will fail if N < npatch.

Anyway, I just ran

  % ABC --run-impls 2 --sim-params 100 0 1
  % gprof ABC >&gprof.out

and gprof has now used up 4 minutes of CPU time and its image size is 1.5 GB. I suspect it will crash soon. 8-) I'm using GNU gprof 2.10.91 as well. I did compile with KCC and full optimization - does gprof now work in that situation?

Oh, yes, gprof finally died after 10+ minutes of CPU time - "Cannot allocate 3087007742 bytes...".

  Jim

> -----Original Message-----
> From: Gabriel Dos Reis [mailto:address@hidden]
> Sent: Thursday, August 23, 2001 6:47 PM
> To: address@hidden
> Subject: Re: [pooma-dev] Profiling POOMA: How to? (1/3)
>
>
>