Hasso-Plattner-Institut Potsdam Operating Systems and Middleware Group at HPI University of Potsdam, Germany
Operating Systems and Middleware Group at HPI

Master level course module on multiprocessor scheduling

Professor Lars Lundberg
Blekinge Institute of Technology, Ronneby, Sweden

Course Overview

The course module consists of two major parts:

  1. Multiprocessor scheduling for high performance supercomputing, 4 hours
  2. Multiprocessor scheduling for real-time systems with guaranteed worst-case performance, 4 hours

The students are expected to prepare themselves by reading some material (to be defined later). Reading this material will take another 8 hours, i.e. for the student this course module requires 16 hours work.

Part 1: Multiprocessor scheduling for high performance supercomputing

In this case there is one parallel program, consisting of a number of communicating processes. The parallel program is executed on a multiprocessor computer. The important performance criterion in this case is the completion of the entire program. Typical examples of such applications include global weather simulation, solving large systems of equations and image rendering in animated films. The way the processes in such a program are scheduled to the processors in a multiprocessor affects the completion time. Ideally we would like to find a schedule that minimizes the completion time of the parallel program, but this problem is unfortunately NP-hard so we need to resort to different heuristic approaches. It is often difficult to know how close to the optimal case these heuristic approaches actually are.

In this part of the course module we explain the supercomputing multiprocessor scheduling problem and explain the difference between static scheduling, where a process is always executed by the same processor, and dynamic scheduling, where processes may migrate from one processor to another at run-time. We will also discuss various heuristic approaches, and finally we will introduce some theoretical performance bounds can make it possible to determine how close to the optimal result a certain heuristic schedule is.  

Part 2: Multiprocessor scheduling for real-time systems

In this case there is a set of independent tasks that are scheduled on a multiprocessor. The important performance criterion here is the response time of each task, and in most cases the most important issue is to guarantee that the response time of a task will even during the worst-case conditions be within certain specified limited, i.e. we want to guarantee the worst-case behavior. We consider two kinds of multiprocessor systems: systems with static scheduling, i.e. systems where a task is always executed by the same processor, and systems with dynamic scheduling, i.e. systems where a task may be executed by different processors at different points in time. We also consider two kinds of task sets: periodic task sets where each task is repeated with a certain frequency (e.g. sampling tasks), and sporadic task sets where each task is only executed once. It turns out that it is NP-hard to find optimal schedules for most interesting cases.

In this part of the course module we explain the real-time multiprocessor scheduling problem and explain the difference between static and dynamic scheduling and the difference between periodic and sporadic tasks. We will also discuss various heuristic scheduling techniques and some theoretical bounds on the worst-case performance of these techniques

Time & Location

  • The course will be given in B.E.2 (Bibliothek)
  • Monday       19.2.2007 11.00 - 12.30 & 13.30 - 15.00
    Tuesday       20.2.2007 11.00 - 12.30 & 13.30 - 15.00
    Wednesday  21.2.2007 11.00 - 12.30 

Slides

Articles

  • Additional articles can be found here.

Leistungserfassung

  • Der Kurs besteht aus drei Komponenten, und zwar:
    - einer Blockvorlesung in der Woche (19.2 - 23.2)
    - Literaturstudium zu ausgewählten Themen aus der Vorlesung (16h)
    - einem Vortrag der Teilnehmer in einem zu den studierten Materialien (6h)

  • In den Leistungserfassungsprozess gehen neben dem Vortrag die zugehörige Ausarbeitung ein.

  • Die Vorträge, deren Themen im Laufe der Vorlesung vergeben werden, sollen im Laufe des SS2006
    als Blockseminar (voraussichtlich Mitte Mai) stattfinden.

  • Für die Lehrveranstaltung gibt es insgesamt 3 benotete Leistungspunkte.

  • Die Belegungsfrist endet am 26.4.2007

  • Bitte sucht Euch ein/zwei Paper aus und meldet Euch bei Andreas Rasche an. Der Termin für das Blockseminar wird noch bekannt gegeben.

Vorträge

  • Die Vortäge werden am Dienstag, den 12.6. um 13.00-15.00 Uhr in der Kommunikationszone Haus C 1. Etage gehalten. Die Länge des Vortrags sollte 45 min. nicht überschreiten.
Johannes Nicolai "Integrating list heuristics into genetic algorithms for multiprocessor scheduling"
"An Incremental Genetic Algorithm Approach to Multiprocessor Scheduling"
Steffen Ryll "Analyzing Fixed-Priority Global Multiprozessor Scheduling"
"Fixed-Priority Preemptive Multiprocessor Scheduling: To Partition or not to Partition"