Techniques for Multicore Thermal Management

Field Cady, Bin Fu and Kai Ren

Techniques for Multicore Thermal Management

•Overview and comparison of techniques•Plus determining the critical thread

•DVFS details

•Thread movement

Taxonomy

• Stop & Go vs DVFS– Stop & Go : suspend core operation for 30

millisecs when temperature above threshold– DVFS : dynamic voltage and frequency

scaling, from control theory

• Distributed vs Global– Apply above to all cores or individually– Performance asymmetry : different demands

on different cores

Taxonomy (cont.)

• Migration– Moving threads between cores– Timescale on order of a millisecond, much

slower than DVFS– Migration is “outer loop” or control, riding on

top of DVFS or Stop-Go

• Migrate “critical” thread– Measure criticality with heat sensor– Or with cache misses as a proxy

Aside : Criticality

• In separate paper, Abhishek et. al. defines “critical” as slowest thread

• If we know which is critical:– Task stealing from critical thread– Guide DVFS to prefer critical thread

• Explored proxies

• 13-32% performance boost in task stealing on 32-core machine

Criticality (cont.)

Cache misses an excellent proxy

Donald and Martonosi : comparison of techniques

• Goal : maximize performance subject to temperature constraint

• Measure performance in BIPS and “duty cycle”, i.e. % useful time, scaled for DVFS frequency

• Run on SPEC benchmarks

• Simulated 4-core processor

Results

All normalized to distributed Stop-Go

Stop-Go was terrible!– Why didn’t they try with lower frequency?– Was 30 milliseconds the right time to stop?

They subsequently focus solely on DVFS, even though the hardware is trickier

Migration Policies

Summary & Conclusion

• DVFS far superior to Stop-Go

• Distributed control helps, esp. for Stop-Go

• Migration helps for Stop-Go

• Counter and Sensor-based migration comparable

DVFS

• Dynamic voltage and frequency scaling (per core).

• Dynamic voltage scaling is a power management technique in computer architecture, where the voltage used in a component is increased or decreased

• Dynamic frequency scaling (also known as CPU throttling) is a technique in computer architecture where a processor is run at a less-than-maximum frequency in order to conserve power.

Challenge

• Multiple cores may need to be manipulated simultaneously to control both power and temperature for a CMP chip. Require a Multi-Input-Multi-Output (MIMO) control

• Application software is always designed for single-core processors. Power shifting needed.

• Heterogeneous cores• Workload of a CMP processor is unpredictable

at design time and may vary significantly at runtime

DFVS

Open-Loop Control

P(k+1) = P (k) + A Δ f(k)

Using Feedback (Close-loop)

• Dynamically change matrix A.

Thread Motion: Fine-Grained Power Management for Multi-

Core Systems

• Limitations of DVFS– Coarse grained

• Initiated by OS in milliseconds• Voltage transition delay ~ 10 microseconds• Too slow to respond fine variations in program

behavior (Cache miss ~ nanoseconds)

– Per-core DVFS with multiple VF settings• High cost of off-chip regulators• Bad scalability with a large number of cores

Motivation

• Idea of Thread Motion– Moving threads between cores with two VF domains– Threads experience virtually continuous Voltage

Thread Motion

• TM Manager– A separate embedded microcontroller running TM

algorithm

• Effective IPC

– maintain a table of IPC for each application– high IPC – compute-intensive– low IPC –cache miss, memory access latency

Thread Motion

• Movement Policy– Assign a thread in a compute-intensive phase

to a high VF core– Intra-cluster movement considered first

• Trigger point:– TM-interval : fixed intervals ~ 200 cycles– Miss-driven : move a cache-missed thread

Thread Motion: Algorithm

Thread Motion

BetterQuality