Skip to content

2018 03 13

Jump to bottom
Andre Merzky edited this page Mar 13, 2018 · 1 revision

Agenda

- updates on release process
  - code reviews are voluntary
  - number of branches are to be reduced -> implies prioritization of work
  - more point releases from master (not necessarily from devel)
- updates on overhead measures (time, utilization)
  - use 'utilization' as additional measure (to time) as metric for overheads
  - support in analytics for computation (and some basic plotting)
  - 4 default plots for individual sessions
    - events over time (scatter)
    - core usage over time (scatter / stacked bar ?)
    - concurrency over all stages over time
    - overheads over time
- planning for RCT Classes [ii]
  - data transfer scheduler - Matteo
    - rather data aware scheduler
  - scheduling is Knapsack - find something better? - Andre
    - different agent scheduler base
  - node-local storage - Vivek
    - needs transfer measurement
    - use cases exist
    - not to compete with MPI-IO
  - handle straggler ??? - Iannis, Srinivas
    - use case not completely convincing (AM)
    - may generalize to performance aware scheduler
  - MongoDB --- - Jumana
    - decouple DB (store unit info) from notifications (2-way comm channel)
    - RP uses ZMQ internally -> extend to UMGR/Agent comm?
    - out of scope: not on component level 
  - Dask for scheduling RP CUs? - George, Will
    - needs isolation and profiling
  - George
    - chunk data for compute units
    - leaflet finder

  - data aware scheduler          ok
  - scheduling is Knapsack        ok
  - node-local storage            ok
  - chunk data for compute units  ok
  - MongoDB                       --
  - handle straggler              ??
  - Dask for scheduling RP CUs    ??
Clone this wiki locally