Howdy EBers,
re: toolchains
First, let me state that I am most appreciative
of the EB people who make things happen and guide
the rest of us. I am so very grateful to have a project
like EasyBuild and even more grateful to the people that
keep it going day after to day. And I promise I will try
harder to contribute more (and learn git) once we get
a hold of the 5 clusters (x86, p7 HPC, p7 BigData, BG/Q
and cloud) that were suddenly dumped on our porch
last year... without any additional personnel to
make them go. Without EB, we'd have a much bigger
mess on our hands.
However, that being said....
We (EB) really need to come up with a more timely manner of
declaring toolchains.
It's interesting... while reading this morning's mail (probably
while the telecon was going on across the pond), I noticed
a notification from Intel for a new MPI release.
The immediate question that arose was "oh shit, now WTF do
I call a toolchain with this change?".
To which, there is no clear answer.
The old toolchain "heuristic" (of sorts), implied maybe
bumping a minor number. e.g ictce-7.1.2 might become
ictce-7.2.2. Similar heuristics apply to the goolf chain
when a new version of OpenMPI shows up.
But that was never clear and I'm glad that EB
has moved toward a YEARrelease numbering scheme.
Two schemes for 'release' have been attempted, both
with merits and problems.
- using letters (perhaps ABCD for 1st-4th quarter of a year)
- using the month (e.g. intel-2015.02)
Also entwined in the issue is the preferred GCC. At present,
as I see it, there are three releases that should be supported:
- 4.8.x - a very stable, time-tested version that is know to behave
well with HPC
- 4.9.x - a less stable version for HPC but yet has important C++
extensions to support newer C++ standards, and
- 5.x - an untested, bleeding edge update that has all kinds of promise,
but as yet isn't ready for prime-time
Now then, here at TAMU, I've given up on ictce... I really
don't see a need to support a compiler based on ancient
GCCs (4.4.7 on RHEL6 and 4.1.2 on our old RHEL5 cluster)
Here at TAMU, I currently concentrate on the following:
goolf-1.7.20 - GCC-484/OpenMPI-184/OpenBLAS-0213/FFTW-334/ScaLAPACK-202
foss-2015a - GCC-492/same as goolf-1.7.20
intel-2015a - GCC-492/iccifort-2015.1.133/impi-5.0.2.044/imkl-11.2.1.133
intel-2015A - GCC-484/same as 2015a
intel-2015B - GCC-484/iccifort-2015.3.187/impi-5.0.3.048/imkl-11.2.3.187
Note: intel-2015b, if I decide to use it, will likely be 2015B with GCC
4.9.2 or 4.9.3.
Of course, with Intel's new MPI this morning, I need something new.
And the intro of gcc 5 just adds to the chaos.
I don't know where I'm going with this. Pardon the rant if that's how
it comes off. I'm just slightly frustrated/disappointed that NO decision
was made this morning on the topic. Personally, I would've preferred
y'all said something like "fuck it, these are the versions our there
today and
are most likely to work... lets put them together and call them
xyzzy-20150708"
and be done with it. Then at least there would be _something_ defined
in terms
of newer tools. Where's the love for GCC 4.8.5 and 4.9.3??? I could
give a rat's
ass about the 5.x series at this point.
I don't have a solution. Wish I did. I'm still trying to figure out
what to call a toolchain with GCC/MPICH2/OpenBLASS/FFTW/ScaLAPACK
for use on the BG/Q (its partial to MPICH2).
But _something_ needs to be done to make sure that toolchains
in EB can keep up with compiler/MPI/mathlib releases.
Sorry if I seem unappreciative. I'm not. I just wish y'all had
all the answers for me before I had the question. :)
Keep up the good work and I promise at least three pints
(and maybe dinner at The Roaring Fork) to all that show up for
the hackathon @ TACC in November.
Jack Perdue
Lead Systems Administrator
High Performance Research Computing
TAMU Division of Research
[email protected] http://sc.tamu.edu
SC Helpdesk: [email protected]
On 07/08/2015 09:10 AM, Kenneth Hoste wrote:
Notes are available at
https://github.com/hpcugent/easybuild/wiki/Conference-call-notes-20150708
On 08/07/15 00:10, Kenneth Hoste wrote:
Hi EasyBuilders,
The next EasyBuild conf call is planned for Wednesday July 8th 2015,
3pm CET.
Topics I have in mind include:
* 2015b common toolchains: GCC 5.1 or 5.2?
* making --robot aware of subtoolchains (update by Alan)
* outlook to EasyBuild v2.2.0 (Kenneth)
Suggestions for additional topics are welcome.
More information about the EasyBuild conference calls is available at
https://github.com/hpcugent/easybuild/wiki/Conference-calls.
Please let me know if you're planning to attend.
regards,
Kenneth
