Commit c397c0bb authored by Victor's avatar Victor
Browse files

F2F

parent 978d1f9d
......@@ -11,16 +11,12 @@ EUABS all hands presentation
- Porting to KNL + perf and energy analysis of the HORSE+MaPHyS+PaStiX stack
A project with tight deadlines
- Wrapup machine access dates planned vs actual
- Draft for the deliverable planned to be submited on 13th of December
4IP project, PMO side
- milestone issued (and updated)
- deliverable work in progress: reflexion/presentation at session (UEABS only)
- timeline
- Wrapup machine access dates planned vs actual
UEABS on PCP machines:
- plan code/machines BCOs
MISSING check the fucking table !!!
PCP-GPU
PCP-KNL
......@@ -44,7 +40,7 @@ deliverable: comments/questions
- what's new
- few comments on perf/energy/scalability
- wrap-up table
- INRIA
- HORSE+MaPHyS+PaStiX stack
- [...]
- Conclusion
......@@ -82,3 +78,52 @@ GPU
UEABS Wrapup Session
********************
missing KNL runs: Specfem some further runs needed, Specfem3D_Globe Code Saturn, riccard (ALYA), jacob (QCD), martti (GPAW), arno (CP2K & NEMO), Mariuscz (PFARM)
possibility to run on frioul (almost same environment, home filesystem shared, account activation form PCP one shoud be really easy)
more and updated informations/momtelcon on the gitlab website
next step is the deliverable -> worked on deadlines for contribution
Talked quiet a bit about that. Lots about figures whould be include and the way they are presented.
BCOs presentation vampirised the UEABS session
good presentation, thank you to all speakers. Please post your presentation to BSCW space:
thanks to every one that participated
EUABS session Mom
*****************
correct slide EUABS/UEABS
Deliverable
time & energy to solution + one column on "performance"
update which metric should be taken from energy tool on website
Showing total energy only is sad because PCP system have been thinked more featured. Whatever benchmarking is showing simple figures. hard to analyse metrics from multiple machine/energy stack.
walltime vs node consumption not good !! GFLOPs/energy max energy envelope
walltime not good because not visible -> use speedup normalize to 1.
update machine specs
Residency not inside the cost figure
QE:
- not the 2nd UEABS testcase
Speedup vs cost pretty interresting plots
PFARM
metrics showed are for same testcase right ? just the number of node changing ? you should present speedup too
QCD
KNL numbind -> cache mode very impacting.
Best Practise Guide session
===========================
Looking for volonteers
latest guides updates
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment