Kernel mechanisms with dynamic task-aware scheduling to reduce resource contention in NUMA multi-core systems

摘要

In NUMA multi-core systems, processors may share different levels of system resources such as the bus, the memory or the cache. When processors attempt to access shared resources simultaneously, resource contention leads to a decrease in performance. This paper proposes a dynamic task-aware scheduling mechanism that reduces resource contention in NUMA multi-core systems. To avoid resource contention, processor cores that share resources are arranged to run complementary tasks. Tasks are classified as either compute-bound or memory-bound, during the run time. Processors are also classified as either compute-bound or memory-bound for running different types of tasks. The proposed dynamic task-aware scheduler then dispatches tasks to run on suitable processors, according to their current types. If power is to be conserved, the processors that are responsible for running memory-bound tasks can be set to the lowest processor frequency to reduce power consumption. Processors can also be set to the highest processor frequency for running compute-bound tasks efficiently. These mechanisms are implemented in the Linux kernel. The experiment results demonstrate that the proposed dynamic task-aware scheduling mechanism improves system performance and reduces power consumption by reducing resource contention among processor cores.