| Table of Contents | ||
|---|---|---|
|
Attendees
Larry - Cavium
Kangkang - Huawei
Tianghong - Huawei
Thomas - ARM
Sean - Qualcomm
Takeharu - Fujitsu
Scott Hare - Qualcomm
David Brash - ARM
Renato - Linaro
Gema Gomez - Linaro
Graeme G - Linaro
Ganesh Raju - Linaro
Anoop S - Linaro
Kanta Vekaria- LInaro
Mike Holmes - Linaro
Francois - Linaro
Martin - Linaro
Agenda
- Technical discussion on Tooling
View file name HPC TCWG BUD17.pdf height 250
Minutes
Overall topics
Kanta: Waiting for ARM to release spec before we can schedule tasks for optimising SVE.
Renato: For now, Linaro's SVE plan is to just “review patches when they come”. But there is lot to be done before even looking at SVE elsewhere.
Larry: Enough low hanging fruit in HPC now, SVE is not a priority.
Kanta: how important is SVE enablement for QEMU?
Fujitsu: Important for us.
Tianghong: Enabling in OpenHPC a priority!
Kanta: Linaro can do initial analysis
Question: Do we have a plan for multi arch aarch64 microarchitecture support (ie ThunderX vs other Hisilcon)
Renato: asked if this is per system, per cluster or other deciding point. Need to chose how this is transparent for OpenHPC.
Kanta: Key focus area is OpenHPC
From Linaro next step is to:
- Investigate what is in OpenHPC
- Build list of libraries needed
TCWG
Initial focusing on LLVM Compiler:
- (Loop optimization & OpenMP)
- FASTMath is not the same as in LLVM vs GCC
- We need some kind of benchmarks for HPC
- Doing single and double precision, not worrying about half precision (but should be fine)
- Starting to make sure all the libraries are available into OpenHPC (libc++, libomp, etc)
Talking to Intel on what is missing in OpenMP and talking to ARM about performance
- Libomp Missing component for OpenMP 4.0 is Offloading, Intel working on it
Larry: Is there any defined process for OpenHPC. Need Eric’s help
Kanta: Not currently and working with their Steering committeeEric is on the TSC and I've been attending the SC calls.
With SVE will collaborate with ARM and upstream libraries.
Make sure they are complete for LLVM/GCC.
It will take a few months to get there, hopefully for GCC 8 / LLVM 5
Looking at SVE support in GDB.
GCC engineers are also working LLVM and optimisations are made in both at the same time.
Resources: With past experience, proposal is to work with member engineers and technically lead them
Others
Tianghong: Do we have a plan for “”open community runtime””, It’s for application, inclu OSN. And plan for benchmark (European Conference on Applications of Surface and Interface Analysis (ECASIA)).
Kanta: Maybe in the future we should look at the OCR efforts and compare with OpenHPC and see if it can be combined.
Benchmarking - UEABS
Action: Eric V(ARM) to highlight what benchmarks are in openHPC
TCWG: Benchmark shows libomp performances better than libgomp but libgomp is more compliance and libgomp comes with GCC.
If the libraries are not available use libgomp and if performance is needed move to libomp.
Will start working on this, once received feedback from HPC, move all LLVM libraries to OpenHPC
Martin: Priority to start with Overlay, PPA
Actions
Re-evaluate OpenHPC in terms of efforts, missing pieces, roadmap
Benchmarking : Need expertise from the people who are doing this
Benchmarking sIngle or multi node
Understand workloads and parameters