Hello all:
Is anyone having problems accessing the LAMMPS source from the Temple U git repository? I'm currently travelling in Japan and was wondering if it is a problem with my local connection.
Axel:
I want to test out the new OMP optimizations. The machine I'm using is a shared memory altix (with 64 processors). Have you ever experimented with this architecture?
Best,
Tamara
dear tamara,
Hello all:
Is anyone having problems accessing the LAMMPS source from the Temple U git repository? I'm currently travelling in Japan and was wondering if it is a problem with my local connection.
at the moment the git protocol access is blocked at temple's central firewall.
it took some convincing to get it opened, but since a network SNAFU a few
weeks ago, the firewall seems to have been restored to the previous
configuration
and my requests to have it opened again have so far been ignored.
that aside, access via http protocol is working fine, but requires that you
have a git client with curl support compiled in. if you have an
existing checkout,
you can change the protocol by editing .git/config and replace git://
with http://
i apologize for the inconvenience, but folks here are not used to our (my) style
of working (we are here only for almost a year) and things always take a little
longer. it'll get batter over time, when people put more trust in my say-so. 
Axel:
I want to test out the new OMP optimizations. The machine I'm using is a shared memory altix (with 64 processors). Have you ever experimented with this architecture?
yes, despite what system administrators tell you, those machines actually
make _excellent_ MPI machines. it will be harder to see a benefit from using
OpenMP (i attribute the myth of having to use multi-threading on shared memory
machines to that fact that you _have_ to have a shared memory machine for
multi-threading, while MPI works both with shared and distributed memory).
the reason for this lies in the NUMA (non-uniform memory access) architecture.
the "shared" memory is only emulated and there is "close" and "far" memory.
the thing that helps MPI scaling is that you have to have high bandwidth and
low latency to emulate shared memory. there may still be a significant benefit
from the /omp pair styles, if you have itanium family processors in
that machine.
itanium processors are _very_ sensitive to "if" statements in the
code, i.e. they
don't like them, and the /omp pair styles have optimizations in them, that try
to reduce as many if statements as possible in the innerloops. there may be
some additional benefit if you have multi-core CPUS, but making OpenMP
scale well, depends a lot on cache size. best is to make a series of tests
with different distribution of MPI vs. OpenMP tasks.
cheers,
axel.
I was having git problems too, I think I tried http://, but maybe not in the right way, but I figured the problem was due to my local firewall. Anyway, I don’t care to fix the problem right now as I’ll only want to get the updated lammps-icms code occasionally.
I just want to point out that you can get the code from going to: http://git.icms.temple.edu/git/?p=lammps-icms.git;a=shortlog
And clicking on one of the ‘snapshot’ links (the top is the most recent).
I didn’t see a (direct) link to this from the lammps-icms page, but it would be convenient unless you are hoping for more experienced people to test your code and to weed out people like me who didn’t previously know about git.
Lisa
I was having git problems too, I think I tried http://, but maybe not in the
right way, but I figured the problem was due to my local firewall. Anyway,
I don’t care to fix the problem right now as I’ll only want to get the
updated lammps-icms code occasionally.
most national labs (and big companies) have two-way firewalls,
that filter not only incoming, but also outgoing traffic. in those environments,
using the http:// protocol is the only option (and accessing the git
repositories) is the only option.
I just want to point out that you can get the code from going to:
http://git.icms.temple.edu/git/?p=lammps-icms.git;a=shortlog
And clicking on one of the ‘snapshot’ links (the top is the most recent).
if you can access the webpage, you can also access the git
repository via http:// this is both using the exact same access
method and would be affected by a firewall in the same way.
one problem with git might be, that it may not include client
level support for http:// due to being compiled without the libcurl
development package being installed when it was compiled.
I didn’t see a (direct) link to this from the lammps-icms page, but it would
be convenient unless you are hoping for more experienced people to test your
code and to weed out people like me who didn’t previously know about git.
thanks for the suggestion. i appreciate it and will look into what could
be done to offer a "get the latest snapshot" option on the homepage.
as you correctly assumed, initially it didn't matter that much, because the
OpenMP changes were fairly experimental and thus having an "only
people that can figure out how to use git"-filter restrict the number of
users would access the code was intentional.
now, that the code has stabilized, matured, and tested in many ways,
it might be a good time to lower the barrier and make it easier for
people to give the code a try. there should be a benefit for almost
anybody that is running on a multi-core machine, so there is no more
reason to make it needlessly difficult. my long-term goal is to have
those /omp versions at some point replace the regular variants in
the mainline lammps distribution. the more people test it and give
feedback, the better. to reiterate, the /omp pair styles can be used
without compiling for OpenMP and still should be a bit faster than
the regular variants (they have optimizations similar to the /opt
package, but not as aggressive)
nevertheless i would prefer to have people use git to access the
code, as that would make is _much_ easier to update. it is very
difficult for me, who has been using source code management
systems for over 15 years, to assess what the potential problems
are. so anybody that is having troubles accessing the git repository,
please let me know and i'll work with you to sort out the problems
and then will post more detailed instructions on the LAMMPS-ICMS
home page to make it easier for the next person.
thanks,
axel.
p.s.: i'd also like to hear what kind of speedup (or slowdown?)
people are seeing and what is the best ratio of threads to MPI
tasks on the individual machines. thanks again.
I just want to point out that you can get the code from going to:
http://git.icms.temple.edu/git/?p=lammps-icms.git;a=shortlog
And clicking on one of the ‘snapshot’ links (the top is the most recent).
it turns out that it can be made even easier to retrieve the latest
snapshot of LAMMPS-ICMS by using this URL:
http://git.icms.temple.edu/git/?p=lammps-icms.git;a=snapshot;sf=tgz
i've added it to the LAMMPS-ICMS homepage next to the instructions
for accessing the git repository.
if anybody has problems with this, please let me know.
thanks,
axel.