Bug 193438 – sci-chemistry/gamess - add ~amd64 keyword

• Status: RESOLVED FIXED
• Product: Gentoo Linux

[Description Alexey Shvetsov 2007-09-22 18:09:22 UTC]

With this ebuild gamess compiles fine and pass all tests on x86_64 


Reproducible: Always

[Comment 1 Alexey Shvetsov 2007-09-22 18:11:51 UTC]

Created attachment 131614 [details, diff]
Patch to support blas on amd64

[Comment 2 Alexey Shvetsov 2007-09-22 18:12:36 UTC]

Created attachment 131615 [details]
new gamess ebuild

[Comment 3 Alexey Shvetsov 2007-09-22 18:23:58 UTC]

Created attachment 131617 [details]
Updated patch

[Comment 4 Jeffrey Gardner (RETIRED) 2007-10-21 16:27:47 UTC]

I'll hook it up when I get some time.

[Comment 5 Markus Dittrich (RETIRED) 2007-10-22 13:05:59 UTC]

(In reply to comment #4)
> I'll hook it up when I get some time.
> 

I suggest that we try using the new pkg-config mechanism
for blas to find the proper libs instead of hard coding the
paths. Also, I don't see any reason why we couldn't link
against the shared libs but I might have overlooked
something.

Markus

[Comment 6 Markus Dittrich (RETIRED) 2007-10-27 20:42:04 UTC]

I've just committed a new revision for gamess with amd64
support. Please give it a shot and let me know should there
be any problems. Everythings works fine on my opteron.

cheers,
Markus

[Comment 7 Alexey Shvetsov 2007-10-29 18:26:29 UTC]

seems to work on my Xeon 5335 cluster
how about using mpi for process comunication =)

[Comment 8 Markus Dittrich (RETIRED) 2007-10-30 12:51:51 UTC]

(In reply to comment #7)
> seems to work on my Xeon 5335 cluster

Great! I guess I can close this bug then.

> how about using mpi for process comunication =)
> 

Sounds great! Want to give it a try and report back ;)
Seriously though, I've had pretty good experience
with gamess' TCP/IP socket code on up to 32 nodes
of a big shared memory machine. If you think MPI
might be more efficient for your system you could 
try PC Gamess which has binaries for several MPI 
flavors I believe.

cheers,
Markus

[Comment 9 Alexey Shvetsov 2007-10-31 18:22:50 UTC]

(In reply to comment #8)
> 
> Sounds great! Want to give it a try and report back ;)
> Seriously though, I've had pretty good experience
> with gamess' TCP/IP socket code on up to 32 nodes
> of a big shared memory machine. If you think MPI
> might be more efficient for your system you could 
> try PC Gamess which has binaries for several MPI 
> flavors I believe.
> 
> cheers,
> Markus
> 

Btw pc-gamess mostly based on gamess source code with some  additions =)
I think that mpi version of gamess will be more sutable for clusters with infinibend interconnects =) Sockets based version works fine up to 64 ccores. But if you try to use more cores than you see that there are some performance ovelap due to ethernet communications

[Comment 10 Markus Dittrich (RETIRED) 2007-11-01 19:03:59 UTC]

Yeah, I can have a look at this at some point time 
permitting. If you would like to take a stab at it I'd
appreciate your comments/patches/...

Thanks,
Markus