QUINN GAUMERECE 259/CPS 221

Improving Performance Isolation on Chip

Multiprocessors via on Operating System Scheduler

Outline

DefinitionsMotivationAlgorithmEvaluation

Definitions

Definitions Fair xyz: xyz under fair cache allocation Slow Schedule: Thread run with high miss rate co-

runners Fast Schedule : Thread run with low miss rate co-

runners

Motivation

Programs running on multiprocessors depend on co-runners

Shared caches aren’t necessarily shared fairly

Does it really matter? If one process suffers then another gains…

Cache Fair Algorithm

Guarantees program runs as fast as it would if resources split equally

Does not actually affect the cache allocationThreads with less cache space will have

lower IPCThreads with IPC higher than Fair should run

for less timeIf a thread’s IPC is lower than its Fair IPC it

should be kept on the processor longer

Cache Fair Algorithm

What does it actually need to do? Maintain approximate fair IPC Keep track of current IPC Scheduler Compensation

Cache Fair Algorithm

Two classes of threads: Cache Fair: threads regulated so that their IPC is

equivalent to their Fair IPC Best Effort: threads where compensatory effects for

Cache Fair threads occur.

Fair IPC Model

For each Cache Fair Thread determine Fair Cache Miss Rate Run with several Co-runners Determine Cache Miss Rates Use Linear Regression

Fair Cache Miss Rate ->Fair IPCDone Online

Implementation

Sampling Run thread with various co-runners Determine Cache Miss Rate for all threads Use Linear Regression to determine Fair IPC

Scheduling Checks IPC and Fair IPC Modifies Cache Fair Thread CPU slice Adjusts Best Effort Thread to compensate

Evaluation

What should our performance metric be? IPC can’t be used, only the scheduler is being changed

Performance variability Difference between running with high and low cache

contention threads

Absolute Performance Difference between normal scheduler and Cache Fair

Scheduler

Program Isolation

Difference between programs run on fast and slow schedule

Cache-Fair is always less than Default

Cache-Fair variability always less than 4%

Absolute Performance

Normalized to fast schedule in Default Scheduler

High IPC programs experience speedup.

Low IPC programs experience slow down

What causes this?Overall absolute

performance is competitive

Aggregate IPC

All programs are run on the slow schedule

When they do not meet their Fair IPC they get compensated

Slow schedule means co-runners utilize more of cache

Side Effects

Best effort threads are also effectedSide effects can be limited by increasing

number of Cache Fair and Best Effort Threads.

Questions?