Supercomputers and Scheduling: Applying a New Tool to an Old Problem

Farideh Samadzadeh, Ph.D.Professor of Computer Science

Oklahoma School of Science and MathematicsOklahoma City, Oklahoma

Supported in part by a grant from

OSSM?

► A public, two-year (Jr/Sr), residential school► Advanced math and science curriculum► Recruits gifted students statewide

► Many students participate in a mentorship program under the direction of scientists and engineers in a variety of fields during their senior year at OSSM.

► Two-year in-house mentorship program in Parallel & Distributed Processing done by five OSSM students.

Where are they now?

• Rayus Kulpicki: Univ. of Tulsa, Computer Science sophomore

• Michael Liang: MIT, Biomedical Engineering sophomore

• Robert Lindsey: RPI, Computer Science sophomore

• Raquel Phillips, Univ. of Tulsa, Computer Science freshman

• Nozlee Samadzadeh, Yale, Computer Science freshman

Why is scheduling important?

• One of the challenges of supercomputing is that although you have a very impressive computing power at your disposal, harnessing this power depends on the programming model used.

• There is always competition for acquiring a large number of processors.

• Minimizing the number of processors needed for maximum speed up is a desirable goal.

Major issues in parallelism?– Program Partitioning: task systems result from

the application of a partitioning scheme to a program. Partitioning criteria depend on the desired level of granularity.

– Scheduling: assignment of tasks to processors.

Multiprocessor Scheduling

• Scheduling: assignment of n tasks to p independent processors.

• The execution sequence of the tasks may be constrained by certain precedence relations among them.

• Objective: an assignment of the tasks to processors, minimizing the overall execution time.

• Two types of tasks systems: independent, dependent.

• Independent Task Systems– No communication or synchronization.– Tasks may be executed on any processor and

in any order.– Main concern: load balancing.

• Dependent Task Systems– Require communication.– May require synchronization.– Cannot be executed in any order.– Main concerns: correct execution sequence

and load balancing.

Scope of the Research

• Developing dynamically- and statically-based scheduling algorithms for:– Independent task systems– Dependent task systems

• Studying the behavior and the performance of the developed algorithms under OSCER’s distributed memory supercomputers Boomer and Topdawg.

• The programming model used was SPMD (Single Program Multiple Data).

• Platforms explored:– PVM under a small in-house distributed

network.– MPI under OSCER’s Boomer and Topdawg. – MPI under a small in-house BCCD network for

local testing and debugging.

Dynamic Scheduling with Vertical Partitioning using Independent Path Scheduler

Static scheduling using

Figure 3: Performance of a typical task system as a function of the required number of processors (regardless of the number of the tasks and their processing times)

Lessons learned?

• Today’s supercomputers are tomorrow’s desk-top computers.

• Early involvement in supercomputing education and application development is the key to training tomorrow’s scientists and engineers.