Good Alaskan Morning,
As a Board member, since the intent of creating the Fedora Grid
project was announced, I've been trying to get an idea of what sort of
open community we can build around it and its concepts. Unlike most
services inside Fedora, the Nightlife grid is not primarily aimed at
Fedora users and contributors, its really aimed at researchers who
have computationally intensive projects to solve. Nightlife turns
Fedora users into resources for researchers. The hard problem is
going to be encouraging researchers who are doing scientific
calculations, who by and large are not Fedora contributors, to place
compatible projects on the grid. Getting Fedora users to run clients
will be far easier a problem to solve in comparison.
So its time for me to ask the really hard question.... what's in it
for the researchers? I've been asking around a little and you have a
fundamental problem. In my very unscientific estimation, small and
mid-scale scientifiic research numerics is dominated by proprietary
tools such as MATLAB and IDL. I'm talking about processes that are
easily divided up for multi-processor calculation, like particle in a
cell simulations, or parameter studies. There's a lot of code hand
built code out there, in a questionably licensed state, which was
never really meant to be distributed or used by other people. A lot
of that code is not something we can run natively on a Fedora system,
nor could we even distribute the tools. This is a problem.
Beyond that, once you get into massively parallelized simulations,
codes still may require proprietary libraries like NAG, or even
proprietary compiler chains. These are the sorts of codes which
expect to be run on large highly parallel cluster. and use openmpi or
some other technology to negotiate disturbingly large matrix inversion
calculations. However, these sorts of codes are more likely to be
organized into an active upstream project with acceptable licensing
for easy distribution and use by researchers. When national labs or
universities invest in HPC linux clusters, these are the sorts of
codes they do it for. I'm not sure these sorts of things are easily
griddable for NIghtlife. And even if they are, the upstream
developers would need to make the invest to make sure these simulation
codebases run on Fedora. The codebases maybe licensed openly, but
they may require proprietary dependencies to fully function. This is a
problem.
So where does that leave us? I think a lot of work will have to be
done to find the first few compelling research project which would be
compatible with Fedora Nightlife. Even if there is a need for spare
Fedora CPU cycles, I'm not sure the need can be met using available
open technology without researchers doing a significantly re-tooling.
So why should researchers re-tool? As a researcher myself, I know my
personal answer to the question. I'm slowly re-tooling the analysis
I do away from the existing trove of IDL routines to python. I could
have just bought another IDL license and used the available storehouse
of IDL tools built up already over the last 10 years before I joined
this project. And if I were compelled to by my employeer I would be
using the IDL. What is going to be 'enough' enticement to compel a
large number of researchers to start doing something similar? Or
better yet, to compel their project or institutional management to
re-tool away from proprietary solutions? I'm not sure the scientific
funding structures encourage individual researchers seek open
solutions as strongly as they could. Will the Nightlife grid itself
be an important enough resource to encourage that re-tooling? I'm not
sure. Its novel, and its cool.. but I'm just not sure there is a pent
up need for spare CPU cycles so severe in the academic research
community that it would cause a migration from proprietary tools.
But I think the Niightlife grid can be a piece of a larger Fedora
effort to establish standard open HPC software stack that encompasses
highly distributed community grid which Nightlife aims to be, but also
deeply parallel, tightly networked clusters which researchers have
access to as part of their grant funding either at national lab
facilities or locally at home institutions. With a fully open HPC
framework that we can ship in Fedora and thus EPEL/RHEL, we then give
simulation codebase projects something to target as a platform. We
might even be able to encourage institutional big iron to participate
as Nightlife nodes by running the open HPC stack that Nightlife is a
part of.
So that's the challenge. For Nightlife to be compelling as a service
to researchers, I think 'we' have to re-invent Fedora as a platform
for scientific HPC development which can compete with the likes of
MATLAB and IDL...even in environments which have institutional
site-licenses for this proprietary tools. We have to make Fedora
compelling for HPC generally. How do we do that? I have no idea.
I'm going to be trying to find the people who do have ideas, the
ability, and the motivation to see things changed. I will be pointing
HPC stakeholders to this list to have the discussion. All I have is
questions, questions like these:
What should Fedora provide as an open functional HPC stack?
Could such open HPC stack be configured out of the box for common usage cases?
What large existing scientific simulation projects are candidates for
inclusion into Fedora?
Are those upstream project developers interested in seeing their
framework included in Fedora?
Can those codebase be adapted to be run on a highly distributed grid
such as Nightlife?
Is it possible for institutional clusters to join the Nightlife grid
as clients?
Why would institutional clusters choos to join the grid?
-jef
