No performance variation with range of Lookahead arguments in Slate PGEQRF

[hut...@illinois.edu]

Hi all ,

I have been playing around with Slate's PGEQRF routine for a while and have noticed no performance/cost change if I modify the Slate::Option::Lookahead argument. From the statistics I generate, it seems as if the underlying schedule is not changing at all.

For example, using 16 nodes of Stampede2 with 64ppn, so 1024 MPI processes, Slate runs in 5.24 seconds regardless of the argument I pass into Slate::Option::Lookahead. I have tried arguments 0,1,2,4,8,16.

Do you know if the Lookahead feature is disabled (perhaps just on non-GPU machines or something), or if there is some other explanation?

Thanks!

Edward Hutter

[hut...@illinois.edu]

I forgot to mention that the matrix I was testing was of size 65536x8192.

[Mark Gates]

Lookahead should be active. We will investigate to see if we can reproduce the issue.

Mark

[hut...@illinois.edu]

I invested a bit further. Lookahead pipelining is active, but does not seem to provide much benefit, at least on Stampede2.