Is Process Scheduling a Dead Subject? Neil Audsley University of - - PowerPoint PPT Presentation

Real-Time Systems Group

Is Process Scheduling a Dead Subject?

Neil Audsley University of York, UK

Tuesday, 5 July 2011

Introduction

q Scheduling theory & research driven (to some extent) by hardware trends

Ø eg. SMP trends have led to increase in multiprocessor scheduling

q So where are hardware trends taking us now?

Tuesday, 5 July 2011

Assumptions

q The schedulable entity is the task q No. of schedulable entities (tasks) >

• no. of CPUs

Ø otherwise scheduling theory largely trivial!

q Architecture is essentially fixed q Memory hierarchy is effectively a single static (uniform)

Ø accessed in a reactive / on-demand manner

Tuesday, 5 July 2011

Observations / Trends

q Parallelism

Ø 4, 16, 64, ..., 1K, ..., 10K CPUs on a single die. Ø “Manhattan” Network on Chip interconnect

q Simpler CPUs

Ø Power, cooling, limited cache coherence etc

RAM Multiple CPUs

Tuesday, 5 July 2011

Observations / Trends

q NUMA (Non-Uniform Memory Architecture)

Ø Non-uniform access times

q Huge bandwith / bottleneck

Ø Use must be smoothed for energy / heat reasons

q Intermediate memories RAM Massive bandwidth Intermediate memories

Tuesday, 5 July 2011

Observations / Trends

q Heterogeneous & Reconfigurable

Ø CPUs (and network) can have dynamic capabilities Ø instruction sets, speeds, architecture (VLIW, GPU) Ø Function accelerators Ø Tasks can have multiple implementations

RAM Variable CPUs Function accelerators

Tuesday, 5 July 2011

Challenges

q More CPUs than tasks

Ø Schedule Data not Tasks Ø Data (and code) movement critical to timeliness Ø Offline issues

– characterising tasks / systems in terms of data movement requirements

q Time-space / heterogeneous

Ø Assigning multiple CPUs (and/or function accelerators) to a task Ø Change the architecture (CPU configuration / function accelerators) Ø Necessarily dynamic / online

Tuesday, 5 July 2011