Distributed Lookup Tables – SIParCS 2009 Talk
I finally got the video of my talk at NCAR summarizing my work over the summer of 2009 in the SIParCS internship program. Check it out!
Distributed Lookup Tables – Alan LaMielle – NCAR SIParCS 2009 from Alan LaMielle on Vimeo.
This is a talk I gave that summarizes my work over the summer of 2009 in the SIParCS program at NCAR in Boulder, Colorado. I looked at reducing the non-scalable memory footprint of a set of lookup tables in an atmospheric model within CCSM—the community climate system model. Our approach was to utilize a feature of MPI 2 called one-sided MPI by distributing the lookup tables across MPI tasks. One-sided MPI allows one to remotely access memory of other MPI tasks.
Great presentation, Alan! I see a Ph.D. in your near future.
Good talk. I was wondering if it would be possible to some sort of primitive compression as well to cut down on memory footprint. Nothing too fancy, you don’t want to cut into performance.
Possibly, yes. As I mentioned, we found that something like 60% of the data wasn’t even being touched in a year. That’s a huge chunk of wasted space and therefore we need to do _something_ about it. Compression may be an option, our approach may be. There are quite a few ways to approach solving the problem. Ours was just one of them. : ) Originally this project was hoping to use some sort of PGAS (Partitioned Global Address Space) language such as Chapel, Co-array Fortran, or UPC. We decided against this approach as integrating these languages into the existing codebase would have been more difficult.
Hey I’m really quite interested in the idea behind this work.
Did you ever write it up / publish it?
If you could drop me an e-mail to discuss it I would really appreciate it.
Olly
We never got to anything beyond a prototype implementation. However, it’s still a very interesting topic to me. I’ll send you an email.