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:
Accelerator Meeting @ 12:15-12:45 PM in 4403 SC
In Attendance: Prof. Kale, Lukasz, Michael, Ronak, Harshit, Eric

Possible Projects (from last notes)
-Machine layer
-Profiling various code for ratio of speed between hardware
-Two big pieces
--Hetero LB
--Multi-PPN
-Optimization/vectorization

-Long term GPU work?
--Memory management (also for Phi?)

Long Term Interests
-Use BW effectively with and without GPUs
-Research interest - hetero LB
--bringing both objects to the equation some better suited to host/accel/both
-Need to get mechanisms in place/working to get there
--For both GPGPUs and Xeon Phi
---Offload API + Dave's Kunzmans work vs gCharm
----Which to follow?
---Xeon Phi - what is our approach?
----Solve (ad-hoc) ppn problem

Memory Bandwidth
-Currently we only look at load on CPU, not MB
-May want to examine separately (per Abhishek's work) and assign separately
-Broader issue, computational load not only LB
--Memory bandwidth and occupancy
--Also network
-Greg Brotevetsky has already had some ideas on this issue about co-scheduling

-Currently PE as core, think about PE as node/SMP domain
--Variety of resources available to it
-Can't currently co-schedule today in charm
-This new scheduler in new paradigm could correctly assign work to correct 
cores
-Need a global scheduler across all cores to direct individual schedulers
--So they can change priorities accordingly (currently static, we need to 
make the queue's dynamic)
-Need orchestrator? Or just have individual cores assign
-May need to add friction terms to dampen oscillation

-Applying control theory to our work (non-linear and discrete)

Projections
-Has some GPUManager integration
-Eric/Harshit may want to work on this to figure out problem from blocking
-Dave may have expanded this, need to investigate further

Benchmarks
-LeanMD has low mem b/w usage, things fit in cache
-Jacobi2d has high mem b/w usage

Work for Next Week
-Harshit - has work
-Eric - review why openatom resulst were bad
-Michael - work on multi-ppn
--Maxed fixed pattern or continugous series
---Recognizbale by regex to easily calculate
---Could just sort them internally beforehand
--Have a limited subset of ppn types (Xe6, xk7)
--Look at common case first and figure out
-Ronak - profiling/measurement/LB
-Lukasz - TRAM

ToDos
-Dynamic priority queues

Future
-Involve Cyril (StarPU)

To view this item, click on or cut-paste
https://charm.cs.illinois.edu/private/tms/listlog.php?param=1490#12622

--
Message generated by TMS