Proposal Slides

Download Report

Transcript Proposal Slides

Migration Cost Aware
Task Scheduling
18-743 Energy Aware Computing
Shraddha Joshi, Brian Osbun
9/24/2013
Outline
• Motivation
• Problem to be solved
• Methodology
• Proposed system configuration
• Simulation environment
• Cost and scheduling formulation
• Milestones
• Q&A
2
Motivation
• Heterogeneous systems can improve power/
performance efficiency by scheduling tasks on the
most suitable cores
• Compute-intensive tasks run on high-performance cores
• Less intensive tasks run on low-power cores
3
Motivation (cont.)
• Dynamic schedulers allow this task mapping to be
updated during program execution
• Migrating threads to new cores can have hidden
costs
• Moving architectural state
• Increase in cache misses
• Congestion on interconnect during transfer
4
Problem to be solved
• Most task schedulers ignore migration overhead
• Solution: quantify and consider the task migration
cost when evaluating scheduling possibilities
5
Methodology (configuration)
• Cluster: a set of cores with different performance
levels
• Shared L1 cache per cluster
• System: a set of multiple clusters
• Shared L2 cache between all clusters
• This creates a cost difference
• Intra-cluster migration
• Inter-cluster migration
6
Methodology (simulator)
• Use the Sniper Multi-Core Simulator
• Interval based, x86 simulation
• Supports heterogeneous configurations
• Python interface for runtime control
• Ability to schedule tasks among cores
• Produces performance and power statistics
• Integrated with McPAT framework
7
Methodology (scheduling)
• Determine the conditions for initiating a task
migration
• Current IPC can be a good predictor for the future IPC
• Determine acceptable ranges of migration costs
• Migration cost related to predicted number of memory
intensive instructions
• Choose whether migration provides a net benefit
• Cluster architecture allows different tiers of migration
8
Milestones
• Be able to quantify migration cost in terms of cycles
• Develop a dynamic task scheduling algorithm
• Incorporate migration cost into our algorithm
9
Q&A
10