Scheduling parallel programs in distributed systems

Thesis/Dissertation · Fri Jan 01 04:00:00 EST 1988

OSTI ID:7251498

Rommel, C G

Scheduling parallel programs under the processor-sharing discipline for uniprocessors, multiprocessors, and distributed systems was studied. Two classes of parallel programs are considered: those without any IPC (called Fork-Join jobs) and those with asynchronous and uniform IPC (called clusters). The study is divided into two parts: (1) develops analytical solutions for Fork-Join Jobs on uniprocessors and multiprocessors; and (2) develops and evaluates via simulation Fork-Join jobs and clusters on distributed systems. The types of site scheduling studied are TS-PS where tasks of a job are scheduled independently at processor-sharing servers, JS-PS in which tasks of a job are scheduled as a single entity at processor-sharing servers, and FCFS where tasks of a job are scheduled independently by order of arrival. For Poisson job arrivals and exponentially distributed task service times, analytical solutions and computationally efficient bounds were found for Fork-Join TS-PS and JS-PS job response times. An algorithm was developed to schedule parallel programs in distributed systems. Over a wide range of parameters the algorithms was found to be superior to both no-load balancing, NLB, and shortest-queue first scheduling, SQF.

• Research Organization: Massachusetts Univ., Amherst, MA (USA)
• OSTI ID: 7251498
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990200* -- Mathematics & Computers
ALGORITHMS
ARRAY PROCESSORS
DATA PROCESSING
DISTRIBUTED DATA PROCESSING
MATHEMATICAL LOGIC
PARALLEL PROCESSING
PROCESSING
PROGRAMMING
TASK SCHEDULING