1. Store Recycling Function Experimental Results
Motivation
Overview
Verification Run
Verification run
checks correctness of recycling functions on smaller problem instance
Production run
executes task graph with verified recycling function on actual problem instance
Emergence of multi/many-core architectures
Need for effective parallel programming models
Dynamic task graph scheduling
task scheduling via work stealing
memory management
Our approach: recycle memory assigned to data-blocks among tasks via store recycling functions
Recycling constraints
T1 can recycle T2 if both T2 and all of T2’s uses causally precede T1  ensures no premature recycling
Two tasks T1 and T2 can recycle the same task T3 only if T1 can recycle T2 (or vice versa)  ensures no concurrent recycling
Track causality relationships between tasks via vector clocks
Auto exploration of recycling functions
recycling candidates: immediate and transitive predecessors
ask user the dependence structure, then enumerate all traversal paths
Background
Production Run
Representation as a DAG
vertices (tasks)
edges (dependences)
Guarantees
no concurrent recycling is allowed
Memory management
single assignment: likely to run out of memory
garbage-collection: requires use count or last use specification for each data-block
a data-block recycled too early is recomputed through re-execution
Store Recycling Function
Experimental Results
Setup
Intel Xeon Phi: 61 cores (244 threads), 8GB memory
Bench.: Cholesky, FW, Hotspot, LU, Rician, Srad, SW
3) Re-execution Overheads
Maps a task T1 to another task T2 so that output of T1 occupies the same memory as the output of T2
Ex: Recycle(B) = A
Challenges:
characterization of the sufficient conditions for a recycling function to be correct
- Recycle(B) = D
- Recycle(B) = A and Recycle(C) = A
determination of the most memory efficient recycling function given a set of candidates
efficient representation of recycling functions during runtime
guaranteeing correct execution for every possible schedule and problem instance
1) Comparison between single-assignment
and recycling -- Auto recycling overheads
2) Recycling Function
Verification Costs
Incorrect execution
Fig3. Re-execution overheads with a representative incorrect recycling function
Fig1. Cholesky with small (left) and large problem instance (right) for varying number of threads
Fig2. Associated costs at 61 threads
Tab2. Number of recycling functions checked and verified as correct
Fig4. Distribution of incorrect recycling functions to overhead bins
Tab1. Memory consumption in MBs