~/talk

Cost Models for Locality and Parallelism

links:none on file
slides:none on file

Defines a program‑centric cost model that counts transfers on nested parallel DAGs and shows how to map it to a tree of caches using a parallel depth‑first scheduler that pins maximal M‑sized subtasks. Introduces parallel cache complexity (locality) and effective cache complexity (adds imbalance penalties), gives tight bounds for matrix multiply (parallelizability 1.5) and sorting (1.0), and reports a practical result: sorting 100 million doubles in about one second on a 32‑core machine.

AI-written summary · about

  1. 0:00Start
  2. 0:07For scalability...
  3. 2:46This Talk
  4. 3:13Parallel Program Directed Acyclic Graph
  5. 5:32Cost Model for Time
  6. 7:28Upper Bounds (Algorithms)
  7. 8:37Lower Bounds
  8. 9:46Problem: Unwieldy Model
  9. 10:53Separate Cost Model & Schedule
  10. 11:59Program-Centric Cost Models
  11. 13:07Realistic Machine Model: Tree of caches
  12. 14:10Program-Centric Cost Model for Sequential Algorithms Cache Oblivious Framework JFLPR 99
  13. 16:42Program-Centric Metrics for Parallel Programs For nested parallel programs on shared memories.
  14. 17:32Nested Parallel DAGS
  15. 18:32Locality: Parallel Cache Complexity
  16. 20:07Scheduling on Tree of Caches Annotate tasks with size, schedule based on size
  17. 21:10Scheduling: Cost Model to Machines
  18. 21:35Schedule preserves locality Communication Costs
  19. 22:21Effective Cache Complexity BFGS'11 Extend Q* to include cost of imbalance
  20. 26:06Summary Important to quantify locality and parallelism.
~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.