A low cost, high performance dynamic-programming-based adaptive power allocation scheme for many-core architectures in the dark silicon era

Abstract

Abstract-Power consumption of many-core chips increases in such a rapid pace that it will soon exceed the chip's affordable power budget. As a result, design of a many-core chip has to address a significant performance challenge under a tight power budget constraint. This problem becomes more prevalent as more of the frequencies and/or voltages of on-chip resources in a manycore chip can be tuned, where heuristics based power allocation approaches often lead to poor performance. Another important problem is that the input power budget of a many-core chip might actually undergo a rapid change at run time. In this paper, the performance optimization problem is formally formulated, and an Optimal Power Allocation method using Dynamic programming (OPAD) is proposed to solve this problem. OPAD has a linear time complexity, and it is quite scalable to the problem size. Extensive experimental results have confirmed lower application execution time of OPAD than that of other competing power allocation methods, i.e., 20%∼30% reduction in applications' execution time over three competing methods. The runtime and hardware overhead of OPAD are also shown to be very small, making it suitable for adaptive power allocation in future manycore systems.

Dynamic Programming-Based Lifetime Reliability Optimization in Networks-on-Chip

“…The problem can be defined as maximizing performance given fixed lifetime budget. This is similar to the power budgeting problem [29,31] which maximizes performance under limited power budget. However, the lifetime budgeting is different as the aging process is in a long-term scale.…”

An efficient runtime power allocation scheme for many-core systems inspired from auction theory

“…Experiments are performed using an event-driven many-core simulator [29,31]. This cycle-accurate simulator faithfully models the processor microarchitecture, the cache coherence protocol, and the communication infrastructure.…”

“…The last step of frequency scaling using a hardware dynamic programming network also has a low area and power overheads, only accounting for 0.5% of the total network-on-chip [29].…”

“…In the last step, the frequencies of the active tiles are set under the power budget, which can be achieved by the optimal dynamic programming network based method as reported in [29,30].…”

The Role of Self-Awareness and Hierarchical Agents in Resource Management for Many-Core Systems

“…However, in all these static mapping strategies, communication between different tasks is not considered [35]. Heuristic-based approaches are not able to achieve optimal results and the performance of such systems may degrade massively when the number of controllable resources is increasing [79].…”