On 11/01/2018 05:10 AM, Chris Samuel wrote:
On Thursday, 1 November 2018 4:58:30 AM AEDT Prentice Bisbal via Beowulf wrote:
I read a publication from LANL on it's architecture years ago, and I believe
to had to program for all 3 different processors to take advantage of it's
architecture. (If soneone knows for sure, please correct me if I'm wrong.)
I'd say that's even more heterogeneous than what we are seeing today
(CPU + GPU).
I think you're bang on the money, here's a presentation from LANL
about Roadrunner which includes programming on slide 25.
https://www.lanl.gov/conferences/salishan/salishan2007/Roadrunner-Salishan-Ken%20Koch.pdf
All the best,
Chris
Thanks for this. Slides 25 - 31 are all good slides, but I think slide
31, titled "Programming a Hybrid Computer" really hammers the point
home. Fortunately, it's all text, so I can cut-and-paste it's contents
here:
Decomposition of an application for Cell-acceleration:
1. Opteron code
- Runs non-accelerated parts of application
- Participates in usual cluster parallel computations
- Controls and communicates with Cell PPC code for the accelerated
portions
2. Cell PPC code
- Works with Opteron code on accelerated portions of application
- Allocates Cell common memory
- Communicates with Opteron code
- Controls and works with its 8 SPEs
3. Cell SPE code (8-way parallel)
- Runs on each SPE (SPMD) (MPMD also possible)
- Shares Cell common memory with PPC code
- Manages its Local Store (LS), transferring data blocks in/out as
necessary
- Performs vector computations from its LS data
It seems like this would be a nightmare to program for. A few years ago
I worked with faculty member who did some programming for the Roadrunner
architecture, and he confirmed this.
--
Prentice
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf