Computational sprinting on a hardware/software testbed
Computational sprinting on a hardware/software testbed
复制标题
DOI：
10.1145/2451116.2451135
复制
发表时间：
2013-03
期刊：
影响因子：
--
通讯作者：
Arun Raghavan;Laurel Emurian;Lei Shao;M. Papaefthymiou;K. Pipe;T. Wenisch;Milo M. K. Martin
中科院分区：
文献类型：
--
作者：
Arun Raghavan;Laurel Emurian;Lei Shao;M. Papaefthymiou;K. Pipe;T. Wenisch;Milo M. K. Martin
CMOS的缩小趋势已经导致了一个拐点，即热约束（特别是在仅采用被动冷却的移动的设备中）排除了芯片上所有晶体管的持续操作-这种现象称为“暗硅”。“最近的研究提出了计算冲刺--在短时间内超过可持续的热限制--以提高响应能力，以满足许多媒体丰富的交互式移动的应用程序的突发计算需求。计算冲刺通过激活备用核心（并行冲刺）和/或将频率/电压（频率冲刺）提升到远远超过系统可持续冷却能力的功率水平来提高响应能力，依靠热容来缓冲热量。先前的工作通过建模和仿真分析了短跑的可行性。在这项工作中，我们研究冲刺使用硬件/软件测试平台。首先，我们研究了未删节的sprint，其中计算在温度变得至关重要之前完成，显示出6.3倍的响应能力，以及通过比赛到空闲状态而提高6%的能源效率。然后，我们分析截断sprint，其中我们的软件运行时系统必须进行干预，以防止过热节流并行性和频率之前的计算完成。为了避免超额订阅的惩罚（截断并行冲刺后的上下文切换效率低下），我们开发了一个冲刺感知的基于任务的并行运行时。我们发现，最大强度的短跑并不总是最好的，引入短跑节奏的概念，并评估一个自适应的政策选择短跑强度。我们报告了使用相变散热器延长最大冲刺时间的初步结果。最后，我们证明了冲刺和休息的操作制度实际上可以胜过热限制持续执行。
CMOS scaling trends have led to an inflection point where thermal constraints (especially in mobile devices that employ only passive cooling) preclude sustained operation of all transistors on a chip --- a phenomenon called "dark silicon." Recent research proposed computational sprinting --- exceeding sustainable thermal limits for short intervals --- to improve responsiveness in light of the bursty computation demands of many media-rich interactive mobile applications. Computational sprinting improves responsiveness by activating reserve cores (parallel sprinting) and/or boosting frequency/voltage (frequency sprinting) to power levels that far exceed the system's sustainable cooling capabilities, relying on thermal capacitance to buffer heat. Prior work analyzed the feasibility of sprinting through modeling and simulation. In this work, we investigate sprinting using a hardware/software testbed. First, we study unabridged sprints, wherein the computation completes before temperature becomes critical, demonstrating a 6.3x responsiveness gain, and a 6% energy efficiency improvement by racing to idle. We then analyze truncated sprints, wherein our software runtime system must intervene to prevent overheating by throttling parallelism and frequency before the computation is complete. To avoid oversubscription penalties (context switching inefficiencies after a truncated parallel sprint), we develop a sprint-aware task-based parallel runtime. We find that maximal-intensity sprinting is not always best, introduce the concept of sprint pacing, and evaluate an adaptive policy for selecting sprint intensity. We report initial results using a phase change heat sink to extend maximum sprint duration. Finally, we demonstrate that a sprint-and-rest operating regime can actually outperform thermally-limited sustained execution.