[petsc-dev] comments on make streams

[Barry Smith]

On Mar 23, 2014, at 12:29 PM, Jed Brown <jed at jedbrown.org> wrote:

> Handy visualization, though I think the red text is excessive [1]

   Ok, I’ve removed it.

> and as
> something unrelated from correctness of the installation, I'm not sure
> it should be displayed as part of "make all", especially when it doesn't
> run on batch systems.

   In theory it only prints this suggestion on machines without a batch system and it is printed after make test. Not make all.

test:
	-@${OMAKE} PETSC_ARCH=${PETSC_ARCH}  PETSC_DIR=${PETSC_DIR} test_build 2>&1 | tee ./${PETSC_ARCH}/conf/test.log
	- at if [ "${PETSC_WITH_BATCH}" == "" ]; then printf "Now run make streams NPMAX=<number of expect MPI processes you intend to use>\nto evaluate the computer systems you plan use\n”;fi

   Is this not working and it is always printing the suggestion?

> 
> I see a lot of variability between runs (e.g., erratic performance on
> es.mcs.anl.gov attached), so perhaps we should be making box-and-whisker
> plots.  This is a balancing act because we don't want the test to take a
> long time to run.
> 
> While the "ideal speedup" line is visceral, it makes it hard to see
> what's actually happening.  I would either remove that line or add a
> second y axis that is unscaled (second attachment).  Either way, we
> should find a way to report numbers rather than unscaled speedup.

   In my latest push, from this morning, I added the bandwidths on the right axis.

   I agree with Matt I don’t like the other plot overlaid on top. Too difficult to understand. I also don’t like a separate plot, nor do I want to remove the perfect scaling because on a decent system it provides useful information. I have tried the following: the top of the prefect scaling will be removed to make sure that at least 1/3 of the y axis is used to display actual scaling. See attached image (of course it could be changed to use 1/2 the y axis or whatever.) Is this a reasonable compromise.
> 
> It's crazy that a plot of intra-node scalability of STREAMS is not
> posted prominently in computing facility documentation.

  It is disgraceful but not surprising.
> 
> 
> [1] I still disagree with using colored text unconditionally.  I don't
> want control characters in log files and emails.  We could test whether
> output is going to a TTY so long as a separate stream is being sent to
> the screen than the log files.
> 
> 
> <scaling-es.png><scaling.png>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20140323/f36c9b55/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: scaling.png
Type: image/png
Size: 47731 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20140323/f36c9b55/attachment.png>