A Novel Memory-Aware CPU Allocation Policy for Multicore NUMA Architecture

• Published in: Reliable and Autonomous Computational Science pp. 41 - 60
• Main Authors: Kang, Dongwoo, Park, Heekwon, Choi, Jongmoo
• Format: Book Chapter
• Language: English
• Published: Basel Springer Basel 2010
• Series: Autonomic Systems
• Subjects: Local Memory; Memory Intensity; Memory Load; Node Domain; Remote Memory
• ISBN: 9783034800303; 3034800304
• DOI: 10.1007/978-3-0348-0031-0_3

Recent computer systems, ranging from mobile systems to servers, are employing multicore processors such as Intel’s Xeon, AMD’s Opteron and ARM’s Cortex-A9. Also, to reduce bus contention among multiple cores and DRAMs, they adopt the NUMA (Non-Uniform Memory Access) memory organization, wherein each core has direct path to its own local memory, leading to access local memory faster than remote memory. These trends of computer architecture trigger to rethink the internal structures and policies of today’s operating system. In this paper, we design a new memory-aware CPU allocation policy for multicore NUMA architectures that has the following three features. First, it makes a CPU allocation decision based on not only CPU load but also memory load, which enables to decrease the possibility of referencing remote memory. Second, it applies different weight on CPU load and memory load hierarchically and adaptively according to the types of CPU allocation requests. Finally, it utilizes the characteristics of processes such as CPU intensity and memory intensity to accurately estimate the CPU and memory load of each core. Real implementation based experimental results have shown that the proposed memory-aware CPU allocation policy can actually enhance the execution time of applications, compared with the traditional Linux CPU allocation policy.