Ppl-accel mailing list

[Michael Robson <mprobson AT illinois.edu>]

A new log has been added to Task: Accel Minutes by Michael Robson
The text of the log is:
Accel-Node Summit 11 - 1:30 AM in SC 4102
In Attendance: Harshitha, Phil, Eric, Ronak, Michael, Harshit, Sanjay

Single node
- any PE can execute progress fn
- this is a concern at a large number of cores 
  - can check queue without log
  - can't accquire work request without it
- maybe limit to PE0
- another way of enqueue'ing it would be nice

NOTES: Where is gpu manager created?

Accel Framework
- document to write accel programs

Node level 
- w/in node: k's of cores 
- 1 PE per node
  - 1 execution stream
  - endpoint
  - needs a better name
  - OpenMP, OmpSs 
  - Drones & CkLoop
- 1 PE per core
  - Charm++ SMP
  - Charm++ SMP & CkLoop
- Multiple PE's Node
  - CkLoop, OpenMP, OmpSs 
  - Drones & CkLoop

Load balance w/in a node more often (why don't we do this now?) 

Within node LB
- chares are assigned to PE's
- could be re-assigned w/o waiting for global LB
- be able to dynamically change cores we run on
  cores = execution stream
- choose to only run on 1 of 2 SMT 
- How can we enhance current SMP style?

Learn more synchronization primitives

Different units of scheduling and migration

Provide good infrastructure for 3 versions
- should be separated out into different files 
- do not want to parse or anything
- xeon integrated gpu's 
  - same memory domain
    - have shared L4 cache 
    - ~200-900 GFlops
  - OpenCL (maybe OpenMP or OpenACC)
  - Ehsan's work w/ Maria 

Agenda
- M/N model?
  - Also give M = #cores special care
  - Notion of tasks 
    - Final block inside a when of SDAG
  - SMP model w/ transient LB
    - should be adequate
- Use cases 
  - Jacobi
  - LeanMD
- Intersection
  - Keeping data on device and notifying node scheduler
  - OmpSs has a directory which keeps track, what memory where 
    - Do locality based scheduling
  - gOMP or Clang project

OmpSs
- Keeps track of non-ready tasks 
  - and their data dependencies
- Predictive scheduling

Tasks
- Current SMP multi-core
  w/in node LB
  and heirarchical LB
- Hetereogenous LB (Michael + Ronak)
- Multi-target/version scheduling (Ronak + Michael)
- Transient in-node migration (Harshitha)
  - Migrating chares as needed (merge?)
- Refactoring CmiMyRank == LastThread throughout runtime (Ronak?)
  - Phil, Harshitha, Ronak, and Sanjay discuss
- Fix drone thread scheduler, touches previous issue; not production 
(Harshitha)
- M/N scheme
- Shrink/Expand PE/Drone's on node (Harshitha)
  - Relaunch via shrink/expand
- Thread affinity (Harshitha)
- Provide threads/scheduling for on-node model RT libs (OmpSs, etc) (Prateek?)
  - gOMP, Argo
  - really: gOMP on Converse
  - but really really: OpenMP run time on top of Converse
- Sending packing
  - Drone task, CkLoop
- Worker disabling i.e. turn off looking at node queue for execution (Michael)
  - test program
- Task model/API (Harshitha)
- Task queue's (Harshitha)
- SDAG to create tasks
- RO enforcement (Seonmyeong)
- Make OmpSs work well aka experiment & tune w/ on-node model integration 
(Harshitha + Seonmyeong)
  - Charm + X (where X is OmpSs)

Goals
- Hetereogenous LB
- ID threads as: comm, worker, gpu?, io?

Large machines with some nodes w/ gpus and some w/o
- Blue Waters
- LANL ExMatEx?
  - Target LB here would be a good resarch exercise
- Taub and Golub
- LSU SuperMIC

Personal notes:
- Is there a separate scheduler in accel??
  - or does it override the default scheduler?

Final List:
MR/RB - multi-target/version scheduling
RB/MR - hetero LB
HM    - transient in-node migration
RB    - thread creation/ID refactoring
HM    - pe/drone shrink/expand
HM    - thread affinity
HM    - follow up on drone model functionality
        - after ID/creation fixes
PJ    - provide threads/scheduling for on-node model RT libs
MR    - worker disabling node-queue execution
HM    - task API
HM    - task Q
S     - ReadOnly enforcement
HM/S  - experiment & tune w/ on-node integration
        - Charm + X

To view this item, click on or cut-paste
https://charm.cs.illinois.edu/private/tms/listlog.php?param=1490#15725

--
Message generated by TMS