Latency awareness in operating systems for massively parallel processors

点击翻译按钮获取中文摘要

LAOS investigates countermeasures to latencies in the synchronisation of concurrent shared-memory processes. The first stage thereby forms prevention, namlely constructive measures at system-software development time that avert process latencies to evolve at run-time. Where prophylaxis alone cannot help, objective of a second step is to avoid at least strong jitter in those latencies. If even those abatement measures of process-latency avoidance are not effective, attempt of the third stage is to hide latency wherever applicable. Analytical measures that continuously quantify or monitor, respectively, process latencies and try to level and minimise them by means of scheduling are no subject of the project.Emphasis of the project continuation is latency hiding through so called guarded sections. In contrast to the conventional implementation of critical sections, guarded sections are not subject to mutual exclusion of concurrent shared-memory processes. Instead, contending processes leave sequential workably mandates for a process that controls passage through the critical section. The assignments are queued and served following a wait-free synchronised queuing discipline. A mandating process that encounters an inactive guarded section activates it and takes over leadership in queue processing. Contending processes always pass a guarded section non-blocking and wait-free. As needed, they however block due to conditional synchronisation on the event indicating the assignment of a section-distinct computational value to some placeholder variable. In order to support guarded sections, a wait-free synchronised operating-system kernel is developed to provide the required minimal basis of system functions for conditional synchronisation for priority-controlled processes.The project emanates from existing knowledge of protecting critical sections in non-sequential programs in general and operating-system programs in particular. It makes the basic assumption that, through asynchronous sequential execution of critical sections supported by guarded sections, performance of non-sequential programs improves because of increased degree of parallel processes, on the one hand, and strenghtening cache locality of many-core processors in time and space, on the other hand. The specific interrelation between process-scheduling strategies and entry/exit protocols of guarded sections is challenged. Also questioned is the general relationship as to current and legacy systems in terms of applicability, transferability, and generalisation of the developed concepts and techniques.