These kinds of things are rather hard to track in time, because
everything is a moving target (conda and other package managers
constantly get updated, but also version of packages changes), but here
is a bit more details :
- The 10x performance difference was with a user code, which I
unfortunately can't share (nor do I still have a copy of it). It was
about numpy, which may or may not have changed since MKL can now be
shipped with Anaconda.
- FFTW, 2x performance gain : These slides compare between
Conda-provided (and those provided by other package managers) FFTW, and
one which was built on an avx2 cluster, the performance gain is 2x (see
slides 28 and 29 :
https://archive.fosdem.org/2018/schedule/event/installing_software_for_scientists/attachments/slides/2437/export/events/attachments/installing_software_for_scientists/slides/2437/20180204_installing_software_for_scientists.pdf
- Tensorflow, 7x gain for CPU version, slide 28 of this talk :
https://archive.fosdem.org/2018/schedule/event/how_to_make_package_managers_cry/attachments/slides/2297/export/events/attachments/how_to_make_package_managers_cry/slides/2297/how_to_make_package_managers_cry.pdf
This one was not comparing Conda itself, but manylinux python wheels
provided by the Tensorflow team, but no doubt Conda has the same issue
if they build for generic architectures.
Basically, any package that is compiled in a portable manner, such as
what Conda and manylinux wheels do, will have some degree of speedup if
compiled for the target architecture instead. This is typically achieved
by the team of analysts who manage a cluster.
Cheers,
Maxime
On 2018-08-28 2:20 PM, Ashwin Srinath wrote:
I'm very interested to see these examples? We use and advocate the use
of conda environments and I'm happy to be convinced otherwise.
Thanks,
Ashwin
On Tue, Aug 28, 2018 at 2:17 PM, Maxime Boissonneault
<maxime.boissonnea...@calculquebec.ca> wrote:
Regarding performance, we have example of code using Anaconda-provided
packages that run 10 times slower than the same code using locally built
packages, optimized for the cluster architectures. That's not *a bit*
slower, that's a lot slower.
Regarding "cheating on your partner", that analogy is not by me, but the
point he is trying to carry is that Anaconda basically replaces any cluster
provided versions, which HPC center people are working hard to optimize.
Recent versions of Anaconda are even worse, by packaging things like
compilers and linkers, creating conflicts with cluster-provided system
libraries and tools, and creating a lot of debugging problems for users and
support people alike.
Regards,
Maxime
On 2018-08-28 12:48 PM, Rémi Rampin wrote:
2018-08-28 12:27 EDT, Maxime Boissonneault
<maxime.boissonnea...@calculquebec.ca>:
As a side-discussion, I think we should also be wary of using Anaconda,
and tell users not to use it in a cluster environment. For reasons, see
here :
https://twitter.com/mboisso/status/1034476890353020928
Hi Maxime,
All I see in this thread is that "it's like cheating on your partner" (!!!)
and it's "generically optimized software" that might be a bit slower than
locally-built libs (interesting concern when using Python, an interpreted
scripting language (and on the slow side too)).
Could you elaborate on those reasons?
Best
--
Rémi
The Carpentries / discuss / see discussions + participants + delivery
options Permalink
------------------------------------------
The Carpentries: discuss
Permalink:
https://carpentries.topicbox.com/groups/discuss/T1505f74d7f6e32f8-Mad4fadc6a6da6de2b5f2aeb9
Delivery options: https://carpentries.topicbox.com/groups/discuss/subscription
--
---------------------------------
Maxime Boissonneault
Analyste de calcul - Calcul Québec, Université Laval
Président - Comité de coordination du soutien à la recherche de Calcul Québec
Team lead - Research Support National Team, Compute Canada
Instructeur Software Carpentry
Ph. D. en physique
------------------------------------------
The Carpentries: discuss
Permalink:
https://carpentries.topicbox.com/groups/discuss/T1505f74d7f6e32f8-M37b48fabbde91daec4f331d7
Delivery options: https://carpentries.topicbox.com/groups/discuss/subscription
