Best compile and benchmark for haswell and broadwell architectures?

[David Gleich]

We just got a few new large memory machines with Haswell and Broadwell architectures. 

I'm wondering what is the current recommendation for getting Julia to run as fast as possible on these machines.

We have a license for the Intel Compilers and MKL.

So the question is: What's the best way to build from source? 

- Intel compilers?

- GNU compilers? 

- MKL or no MLK?

- Setting for MARCH? 

- Setting for JULIA_CPU_TARGET? Is native the best choice here?

And next, what is the best way to benchmark and test the resulting compile?

It looks like make testall does the testing, but is there a standard benchmark command to confirm we're getting the "fastest" julia? 

I did look around for help in the forms and via google, and found bits and pieces. (e.g. JULIA_CPU_TARGET, for instance...). Just wondering what the latest recommendations and experiences are!

Thanks!

David Gleich

[cormu...@mac.com]

I asked a similar question last week. The standard recommendation is https://github.com/JuliaCI/BaseBenchmarks.jl. It didn't really meet what I was looking for -- a quick, easy to run benchmark to compare machines -- but you may find it useful for your purposes.

[David Gleich]

Thanks for that pointer. I'll take a look!

[Yichao Yu]

On Sat, Oct 1, 2016 at 8:51 PM, David Gleich <dgl...@purdue.edu> wrote:
> Thanks for that pointer. I'll take a look!

I have a collection of really simple ones here[1]. It's not systematic
or comprehensive in any sense but does check a few operations that I
care about (simd and math functions) and was able catch a few
regressions for various reasons before.

[1] https://github.com/yuyichao/explore/blob/6f616e8069728ba749147e3d3f19db511fa97c85/julia/bench_simple/benchmark.jl