Page 1 of 1

Heterogeneous and multi-GPU computing concept of MAGMA?

Posted: Mon Apr 02, 2018 5:05 am
by Klausb

I use an application which solves 4 equations using sparse LA:
Equation A needs more than 200 iterations so it's worthwhile to move it to GPUs (two or more)
Equations B, C and D require 1, sometimes 2 iterations so they should remain on the CPU because the PCIe bottleneck would slow them down.

Currently the a simulation is decomposed using scotch and solved on multiple CPU cores using MPI.

To be able to port the relevant parts to the GPU using MAGMA I need to understand the heterogeneous and multi-GPU computing concept of MAGMA mentioned in the Magma 2.3.0 flyer.

How is a computation distributed across multiple GPUs?

How can I distribute the remaining bits across multiple CPU cores, possibly on multiple sockets?