Experience Using Multiprocessor Systems—A Status Report
ACM Computing Surveys (CSUR)
The impact of distributions and disciplines on multiple processor systems
Communications of the ACM
Computers in Biology and Medicine
| Hi-index | 14.98 |
A special bus structure, the SYNCBUS, is proposed, which supports task scheduling by hardware. It can be used efficiently in multiprocessor-(all processors identical) and polyprocessor-systems (different pools of processors) running real-time multitasking software with a dynamic load distribution. Tasks waiting for execution create prioritized interrupts, which are distributed over all available processors by hardware. An ``optimal'' load distribution, where the tasks of currently highest priority are assigned to the processors of each pool, is automatically set up and updated continually. The SYNCBUS operates independently of the pool sizes; increasing or decreasing the number of processors in case of changing system requirements or faults is possible. Its decentralized structure allows parallel and local handling of task queues and eliminates processor queues. Overhead times are cut down for three reasons: the faster operation of hardware compared to software table searches, avoidance of memory conflicts due to accessing common system tables by multiple processors, and avoidance of scheduling cycles which do not result in an actual reassignment of tasks to processors. A performance analysis using a mathematical model shows that overhead times can be reduced, depending on the system size and the frequency of task switches, by a factor of up to 102.