Summary: | 碩士 === 國立臺灣大學 === 資訊工程學研究所 === 101 === The potential speedup of applications has motivated the hybrid design of
general purpose processors and dedicated accelerators in electronics industry,
but such a design also leads to a serious synchronization problem when shared
resources are considered. This paper explores the joint consideration of task
synchronization and accelerator allocation for real-time systems, and the ob-
jective is to minimize the worst-case total blocking time for accessing share
resources and accelerators of each task so as to meet all real-time require-
ments. For accelerator allocation, we develop an inheritance mechanism of
accelerator privilege, which allows low-priority task to use an accelerator at
the priority of the high-priority task when the high-priority task is blocked
by the low-priority task. Co-working with the proposed accelerator alloca-
tion mechanism, our synchronization protocol efficiently bounds the times
of blocking to 2 for each real-time task. By looking into the currently used
semaphores and the priorities of all tasks offloaded on accelerators, our pro-
tocol is developed to prevent chained blocking among tasks, which could be
caused by most of the existing synchronization protocols that are not aware
of accelerators.
|