[Numpy-discussion] grant proposal for core scientific Python projects (rejected)
ralf.gommers at gmail.com
Sat Apr 20 06:27:02 EDT 2019
On Thu, Apr 18, 2019 at 5:27 PM Marten van Kerkwijk <
m.h.vankerkwijk at gmail.com> wrote:
> Hi Ralf,
Thanks for the feedback Marten, it's very valuable. We've gotten some more
feedback from people experienced with applying for or reviewing NSF (and
DOE, NASA, NIH) grants, it helps a lot in figuring out what to do next.
> I'm sorry to hear the proposal did not pass the first round, but, having
> looked at it briefly (about as much time as I would have spent had I been
> on the panel), I have to admit I am not surprised: it is nice but nice is
> not enough for a competition like this.
> Compared to what will have included some really exciting, novel proposals,
> most damning will likely have been the modest, incremental goals (for a
> large sum of money):
I'm very aware that's how reviewers will look at it. I don't agree it's
true though (not sure if you think that) - the impact on the science NSF
supports of spending on the order of 10 million dollars on the SciPy
ecosystem will be way higher than of building some new facility, or
supporting one more supercomputer, or whatever else will have been proposed.
Another thought, and this does make our job harder, is that it's very
difficult to claim to do really novel things. Because whatever we do must
in the end pass review from and be accepted by the core teams of each
project and the community. Proposing really novel things will mean starting
new projects; which is just a different kind of proposal - more easy to
sell, likely a lot less impactful.
performance improvements, but without any actual sense of what would now
> become solvable (how does it beat throwing more computers at a problem,
> which is cheap?); better implementations of things that exist (arrays with
> units, sparse arrays); better GPU support (feels like something everybody
> and their brother was excited about a decade ago); etc. I also think any
> panel would expect some concrete examples of facilities that would now be
> helped: e.g., how is this going to help LSST analyze its 20TB/night of data?
You're completely right - we should have focused more on this, trying to be
more concrete. Note that it's very hard to come up with provable statements
like "this is what we can do after this proposal that we can't do now", but
we must do better here. If anyone has references that we can use that would
be very helpful; things like the Decadal Survey in astronomy that state
something about the SciPy ecosystem). What we can also do is better
elaborate the impact of not maintaining/evolving our projects.
> Going forward, best may be to explicitly involve the facilities that use
> python - within astronomy, that would include LIGO and LSST,
We did, both of those. We also had a senior LIGO person as co-PI, and
includes a quote from a LIGO spokesperson about this being critical
infrastructure for them.
but certainly also STScI (and other NASA institutes), which actually
> supports SPE already. It would be good especially to show how much money it
> would save them when this is implemented, so that it becomes clear this is
> a net win.
I like that idea.
Indeed, for any future proposal, I'd suggest to involve (or at least ask
> for advice) some more senior people who have been successful before (within
> astronomy, the likes of Steve Kahn, the LSST director; he was at Columbia
> for most of his career, so there is a connection).
We also did that, got advice and a full draft proposal review several times
from a former NSF program manager. As well as from the likes of Fernando
Perez and Brian Granger at the start and Ryan Abernathy at the end.
Your advice about the kinds of people to get involved is all true, but I
don't think this was our main issue because we did all that. Also talking
to the PMs, we know that's critical and did that (the government shutdown
didn't help here though).
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the NumPy-Discussion