On 07/21/14 17:35, Roland Kaufmann wrote:
On 2014-07-22 00:14, wireless wrote:
Once these modules are verified to work, I'll be posting the ebuilds
(gentoo software packages) up on the net for others to enjoy.
Please consider donating them to a public Git repository here, e.g. we
could create opm/opm-gentoo, so that users could add them directly using
something like `layman --add opm
--overlays=https://raw.github.com/opm/opm-gentoo/master/overlay.xml`
Quick answer: Non_problemo.
There is a bit of cleanup, testing and organizational work on the gentoo
modules as they stand before publically releasing the codes (ebuilds).
Even in an Overlay, the Gentoo community has very high (the highest?)
standards related to building up codes. Not only must they work, but we
have to have the "flag options" properly identified and tested. Having a
long relationship to Gentoo, I'm not about to release something that
brings about bitching (scientist types are the worst, which is good)
even in Overlays. I'm guessing a few more weeks, but if "Lady Luck"
finds pleasure in the Gentoo effort, we shall see how quickly we can get
these modules ready. That said, Gentoo is the best distro for building
from 100% sources and optimization. Gentoo also has a very active ARM64
(aarch64) team that surely will be one of the first linux distros to run
on massively parallel ARM64 bit processors, very soon.....
Additionally, from my review and scant understanding of the codes, they
needs to be re-organized such that every module can be tested and (where
applicable) the mathematical techniques underlying the codes need to be
very easy to replace and individually test for performance improvements
and accuracy. Ideally where a choice exists for the various (competing)
math_codes it would be nice if a simple (gui) button could make those
simulation runs easily selectable while ensuring that both data sets
kept for future runs and reference.
If we break out the mathematics, clearly, we can get all
sorts of phenomenal coders to create components to compete on
performance vs accuracy in the model. These underlying math_code
choices needs to be clearly delineated and separated from the model
assumption to enable a collection of simulation runs to be collect up in
preparation for "REAL-TIME" rendering. Whether database support
is necessary needs to be tested and evaluated. Similarly, porting the
various modules that are bottlenecks to run over a distributed
(clustered) file system will afford the opportunity for both massively
parallel approaches as well as using specialty processors, such as GPUs
and FPGAs and arm64.
Arm64 is the dominant processor, because of a paradigm shift in the
fundamentals of processor design. The old constraint that Intel
dominated was based on compaction; that is the simple fact that the best
gains in processor design were do to how densely one could compact the
logic gates. But the shift has resulted because of the fact that the
geometries of the "logic gates" now being sub 10 nm, minimization of the
"heat generation" dominants processor design. So for general purpose
processors, ARM has laid that competition to rest (at least for the
next decade), evidence my the myriad of companies implementing ARM
design for competitive advantage. I look forward to kicking Intel to the
curb. For those protions of the codes requiring SIMD or MIMD, custom,
finely tune algorithms, porting to run on 'bare metal' via either GPUs
or FPGAs is still the current best solution.
Fully implementing competing approachs will allow the simulation runs to
be "component aware" so that repeated simulation runs only have to
"grind" (re_compute) the numbers on data that will change. If/when we
reach this level of intelligence in the simulation architecture, WE can
run simulations in real time (defined here as less than 1 second latency
on the graphical output) while simultaneously making "what if" changes
to the model assumptions and the underlying math_codes used to grind out
the new numbers.
Speed kills; this is a universal accepted concept. If you want OPM to be
"the killer application" (used to benchmark clusters and such) then
every piece of code needs to be "well defined" and modular so it can be
"fair game" for replacement by a better (faster/accurate) performing
component. I hope (and pray) that we are all on the same page? I do
strive to be a "team player" but compromise does not equate to
mediocrity, in my genetics; hence the need to let every piece of code
become "fair game" for improvement.
Is there an "OPM repository" to collect up scholarly publications,
particularly those of mathematical significance, for the convenience of
OUR TEAM ?
sincerely,
James
_______________________________________________
Opm mailing list
[email protected]
http://www.opm-project.org/mailman/listinfo/opm
