Speedup

[Cbarnaud]

Hello All,

I recently made a speedup curve on the supercomputer I use, Myria, and I am astonishing with the results I get.
I linked my results (jpeg format.), I made them with an invariant mesh.
My curve is the way upon the reference but usually, most of the codes are under or just a little upper.
Does anyone try this kind of speedup test? So, is my curve accurate or realistic? (Or it's a "trick" because my number of nodes is too little (max 448)?)
Honestly, I don't know what to think about those results, they are kind of amazing.

Thanks a lot for your answers,
Best regards,
Camille.

[Yvan Fournier]

Hello,

How many cells per core/MPI rank does this correspond to ? There is a "sweet spot" where the mesh size per core becomes small enough so that more things fit in cache, yet there is enough work to do for good scalability. Beyond that, when the data becomes smaller, the MPI latency becomes dominant, and performance goes down.

So the behavior you observe seems quite "normal", though I usually expect a more curved "bump" in the superlinear section than the more regular behavior you seem to observe.

Best regards,

Yvan

[Cbarnaud]

Hello Yvan,

This is the correspondence between cells and processors following the graph:
1 node : 183020 cells/proc
2 nodes: 91510 cells/proc
3 nodes: 61007 cells/proc
4 nodes: 45755 cells/proc
5 nodes: 36604 cells/proc
6 nodes: 30503 cells/proc
7 nodes: 26145 cells/proc
8 nodes: 22877 cells/proc
12 nodes: 15251 cells/proc
16 nodes: 11438 cells/proc

Also, maybe I don't have enough points to get this curved "bump", I'm trying to add more of them in the future.
I'm happy to know that my curve is pretty normal, I could use it later.
Thank you for your answer,
Best regards,
Camille.