Control of Resource-Intensive Computations under Uncertainty. Ⅲ. Dynamic Concurrent Resource Allocation

摘要

Control problems for a heterogeneous system executing heterogeneous search tasks are considered. Data parallelization of tasks under objective uncertainty-for the unknown and rapidly changing flow of incoming requests-is studied. Critical modes of operation under big system workloads are analyzed. The modified sliding scheduling method allows correcting the processing procedure in real time. A way of dynamic construction of a sequence of non-discriminating concurrent allocations of limited different-type resources between users with equal rights is proposed. A concept of normative maximal performance of the system processing one type of tasks in the single-user mode is introduced. Parametric guaranteed estimates are used to decompose tasks with respect to their type. Scheduling procedures according to which tasks that de facto require less computational costs receive the processing priority and are completed first of all are developed.