Sparse Iterative Solvers Scalability Issue in MAGMA Library

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
ayazhassan
Posts: 5
Joined: Mon Nov 03, 2014 8:44 am

Sparse Iterative Solvers Scalability Issue in MAGMA Library

Post by ayazhassan » Sat Dec 20, 2014 6:07 am

I have executed some tests using hepta diagonal sparse matrices on conjugate gradient implementations in MAGMA library. Regarding scalability, we expect execution time, ET(N), to be proportional to O(14N) i.e. the overall number of non-zero elements. Thus, if we use hepta with 2N, we expent ET to grow as O(2*14N) thus just doubling. But the results below for MAGMA CSR shows: N growing from 1 to 30 but ET is growing from 12 to 18 (from 1 to 1.5). Can you please clarify/suggest why the execution time is not scaling as expected?

Image

hartwig anzt
Posts: 90
Joined: Tue Sep 02, 2014 5:44 pm

Re: Sparse Iterative Solvers Scalability Issue in MAGMA Libr

Post by hartwig anzt » Thu Jan 08, 2015 2:38 pm

Dear Ayaz ul Hassan Khan,

Thank you for your question. I assume you refer with "array dimension" to the size of the linear system, i.e. the number of unknowns. Your graph shows data for up to 33.000 unknowns. This is still very small for a GPU-based solver. I.e. the latency of the kernels launched may dominate the overall execution time. Please try with larger systems, when you reach the 1M range, you should indeed observe scaling with 14N.

Thanks, Hartwig

Post Reply