Ppl-accel mailing list

[TMS <tms AT charm.cs.illinois.edu>]

Michael Robson has added a new log to Task Heterogeneity

Log Entry:

Accel Meeting

2:10 - 3:15 in SC 4102

In Attendance: Jaemin, Michael, Ronak, Juan, Seonmyeong, Kavitha, Matthias

Agenda:

1. GPU-Manager
   
   • examples patch
   • stencil2d uploaded
   • in gerrit
   • design for gpu PE
   • +gp, part of +p, etc?
   • specify explicitly
   • atomic queues
   • possible to enqueue directly in GPU PE
   • where should callbacks be executed? on home
2. Load Balancers LeanMD
   
   • LeanMD hybrid API implementation progressing
   • on top of Jaemin's API changes
3. Other Topics
   
   • accel
   • needs refactoring
   • Tracing
   • needs to be added to Jaemin's changes
   • check if it's currently working
   • Bugs
4. Next Two Weeks
   
   • Plans
   • Jaemin
     
     • GPU PE
     • 2 weeks working prototype
     • 1 month full version
     • GPU manager tracing
     • 1 week checked tracing (fixed if easy)
     • 2 weeks updated patch/added to new gpu manager
   • Ronak
     
     • LeanMD implementation
     • 2 weeks finished
     • gerrit review
     • VectorLB for GPU
     • measure GPU load
     • integrate new algo in charm
     • integrate with accel (split execution, etc)
     • Next steps
     • design LB interface for measuring CPU/GPU with hybrid API
     • implement inside API itself
     • integrate in load balancing framework
     • adding new LB strategy
     • NAMD
     • future: reimplement with charm
     • separate target
     • maybe LB target
   • Dong Hun
     
     • cuFFT, waiting on charisma
   • Michael
     
     • charm on gpu as PE
     • 1 week prototype
     • 3 weeks full version
     • not sure where from there
       
       • write threaded methods
       • how to construct tasks
       • look into OpenMP crossover
     • new accel?
   • Todos
   • Lockout chare while executing on GPU
   • Check load balancing while GPU work outstanding
   • look at other papers
5. Charm Scheduler on GPU
   
   • cause a race condition (on global memory)
   • atomic add based barrier
   • does profiler show blocks to SM? should; try next
   • use calls at runtime
   • should it be block or kernel based
   • may never be able to guarantee

View task