Hi
We are considering purchase of a new machine on which we will run ROMS and SWAN. At this stage we just want to buy 1x "reasonably powerful machine" rather than getting into MPI and clusters. Maybe down the track...
The machine we will purchase will definitely have two processors, the question is, for a given budget, is it best to go for blazingly fast single-core processors (lets say 3.6GHz), somewhat less fast dual-core processors (lets say 2.8GHz), or slower still (per core) quad-core processors (lets say 2.0GHz)?
In theory you get more cycles per $ with the quad-cores, however I suspect that many other factors come into play (bus speeds, operating system, compiler, grid tiling, etc, etc) and that the answer may well be model specific.
In any case, does anyone have any experience or even just theoretical considerations they would like to share?
Many thanks
Giles
Single core, dual core, quad core....?
-
timchipman
Hi,
My basic feeling from what I've read in the forums (alas very little real-world experience to back this up)
SMP scaling with ROMS is very very nice (ie approaches linear / ideal). (in contrast to MPI scaling, which is "OK but not as good as SMP" (but since MPI "works better" on "more-affordable-cheaper-to-scale cluster nodes", it is a common implementation/solution)
So, assuming this is true, you would benefit from maximizing your total cpu MHZ within a single SMP machine.
Thus, given (hypothetical numbers), the following systems available at the same cost:
1 x 4.4 ghz cpu == "4.4 ghz total"
2 x 2.4 ghz cpu == "4.8ghz total"
2 x dual-core (four cpu cores total) @ 2.2 ghz == "8.8 ghz total"
2 x quad-core (8 cpu cores total) @ 2.0 ghz 8 == "16 ghz total"
then clearly your quad-core solution is the sweet spot.
(likely in the above example, you will have progressively less ram per core - assuming a fixed budget - but still, you can get fairly affordable 8-core woodcrest / with 16 or even 32gig ram machines from Dell/etc .. so that is somewhat less true now than it was, say, a year ago when ram prices were significantly higher).
Of course, if the scaling curve for SMP (OpenSMP implementation in Roms) is punishing you at only 50-75% of "ideal yield" then maybe the "best bang for your buck" will lay elsewhere. (ie, in dual-cores not quad-cores?)
Ideally, someone with access to
(a) a quad-core woodcrest system, AND
(b) a dual-core woodcrest system, AND Ideally also
(c) a dual-core AMD, plus ideally,dual-core Intel (Coreduo) ..
... and then of course also building the same version of roms on all boxes, and analyze same set of data .. in same manner .. would be able to provide pretty clearly an answer to this question -- at least, for the given test data set
HTH slightly, and sorry I have no real concrete contribution
--Tim
(Footnote - I realize that I do actually have access to ...many .. of systems detailed in the a-b-c list above -- so really ought to do some of those tests myself, and post them back to this thread).
My basic feeling from what I've read in the forums (alas very little real-world experience to back this up)
SMP scaling with ROMS is very very nice (ie approaches linear / ideal). (in contrast to MPI scaling, which is "OK but not as good as SMP" (but since MPI "works better" on "more-affordable-cheaper-to-scale cluster nodes", it is a common implementation/solution)
So, assuming this is true, you would benefit from maximizing your total cpu MHZ within a single SMP machine.
Thus, given (hypothetical numbers), the following systems available at the same cost:
1 x 4.4 ghz cpu == "4.4 ghz total"
2 x 2.4 ghz cpu == "4.8ghz total"
2 x dual-core (four cpu cores total) @ 2.2 ghz == "8.8 ghz total"
2 x quad-core (8 cpu cores total) @ 2.0 ghz 8 == "16 ghz total"
then clearly your quad-core solution is the sweet spot.
(likely in the above example, you will have progressively less ram per core - assuming a fixed budget - but still, you can get fairly affordable 8-core woodcrest / with 16 or even 32gig ram machines from Dell/etc .. so that is somewhat less true now than it was, say, a year ago when ram prices were significantly higher).
Of course, if the scaling curve for SMP (OpenSMP implementation in Roms) is punishing you at only 50-75% of "ideal yield" then maybe the "best bang for your buck" will lay elsewhere. (ie, in dual-cores not quad-cores?)
Ideally, someone with access to
(a) a quad-core woodcrest system, AND
(b) a dual-core woodcrest system, AND Ideally also
(c) a dual-core AMD, plus ideally,dual-core Intel (Coreduo) ..
... and then of course also building the same version of roms on all boxes, and analyze same set of data .. in same manner .. would be able to provide pretty clearly an answer to this question -- at least, for the given test data set
HTH slightly, and sorry I have no real concrete contribution
--Tim
(Footnote - I realize that I do actually have access to ...many .. of systems detailed in the a-b-c list above -- so really ought to do some of those tests myself, and post them back to this thread).
-
glesser
Hi Tim
Thanks for your thoughts.
We followed pretty much the same logic as you laid out and purchased a dual quad-core machine (IBM 3650 rack server with 2.0GHz Xeons)
Runs SWAN beautifully fast under Linux using OMP option. Zero hassle, just installed Intel compilers, built and ran. Nice!
Haven't got around to running ROMS on it yet, but presume that when I do, and I re-read the benchmarking page, then I'll figure out some standard benchmarks to run and post the results for comparison with others...
Cheers
Giles
Thanks for your thoughts.
We followed pretty much the same logic as you laid out and purchased a dual quad-core machine (IBM 3650 rack server with 2.0GHz Xeons)
Runs SWAN beautifully fast under Linux using OMP option. Zero hassle, just installed Intel compilers, built and ran. Nice!
Haven't got around to running ROMS on it yet, but presume that when I do, and I re-read the benchmarking page, then I'll figure out some standard benchmarks to run and post the results for comparison with others...
Cheers
Giles