2016-11-08 HPC Meeting Minutes

Attendance

Committee Members 

NamePresent
Kanta Vekaria (OCTO, Linaro)(tick)
Martin Stadtler (Director of LEG, Linaro)(tick)

Darren Cepulis (ARM)  

(tick)  

Eric Van Hensbergen (ARM)(tick)      

Kangkang Shen (HiSilicon)

(error) 

Bhalchandra Deshpande(Broadcom)

(error) 

Larry Wikelius (Cavium)

(error)

Gary Yurcak (Qualcomm) 

(error) 

Rammohan Peddibhotla (Qualcomm)

(error) 

Grant Likely (HPE)

(tick) 

Jon Masters (RedHat)

(error) 

Elsie Wahlig (Qualcomm)

(tick)  

Steve Heist (Qualcomm)(tick) 
Takeharu Kato (Fujitsu)(tick) 

 

Guests 

NamePresent

David Rusling (CTO, Linaro) 

(error)

Mark Orvek (EVP, Linaro Engineering)  (error)
Andrea Gallo (VP of Segment Groups, Linaro)

(error) 

Anoop Saxena (Project Manager, Segments)

(tick)  

Francoise Ozog (Director of LNG, Linaro) 

(error) 

Leif

(tick)  

Ard

(tick)  

Steve Capper(tick)
Tom Gall(tick)  

Agenda 

  • Previous Actions


  • Discussion on Numa for HPC (Leif, Graeme, Ard)
  • Discussion on microarchitecture detection (Leif, Graeme, Ard)
  • Agree on the time that this meeting should take place
  • HPC collaborate pages (Kanta)
  • Next meeting 22 Nov agenda
    • Engineering status
    • post SC16 feedback

Minutes

  • Previous Actions

    • Martin: No update yet

    • Kanta: Pitch deck for TAB for SC16

  • Discussion on Numa for HPC (Steve Capper, Leif)

Leif discussing:  NUMA basic exists, if we want to have proper NUMA we need to do more work.  Leif suggests that if we want this to happen, we need to get the ARM teams doing scheduling and power scheduling.

There is a pending review for (Steve can you add that proposal  ASWGxxxxx)

Steve: Focussing on topology

  • Discussion on microarchitecture detection (Leif, Graeme, Ard)

    • Benchmarking

    • Link time optimisation needs reliable information from the kernel

    • How many variants?

      • 12 right now could be higher

    • How much of the software will benefit?

      • Can target a certain number of libraries such as BLAS libraries

    • Explore dynamically determining heterogeneous environments

      • E.g little, medium and big cores needs migration within cores

      • Training is needed to optimise against a specific arch

        • Can use 1 set for point training

    • Elsie: Are we thinking about adding GPGPU as part of the heterogeneous arch?

      • Eric: was only focusing on cores but agrees it’s wider

      • Grant: Agrees with Elsie but we need to maintain focus on key areas

      • Key priorities: HPC workloads working optimally for different microarch. Need to looking in to openCL

      • Gen-Z interest

    • Grant: Getting 52 bit address enablement is important. Non trivial amount of work and needed for the distros

    • Distros enablement

    • Next steps: Kernel training methodology using BLAS FFTW test cases

      • Pick a library for self training

  • Agree on the time that this meeting should take place

    • No objection from attendees to move this to an hour earlier

  • HPC collaborate pages (Kanta) - postponed to next meeting

  • Next meeting 22 Nov agenda

    • Engineering status

    • post SC16 feedback