RedCooper: Hardware Sensor Enabled Variability Software Testbed for Lifetime Energy Constrained Application
Yuvraj Agarwal, Alex Bishop, Tuck-Boon Chan, Matt Fotjik, Puneet Gupta, Andrew B. Kahng, Liangzhen Lai, Paul D. Martin, Mani B. Srivastava, Dennis Sylvester, Lucas Wanner, Bing Zhang · eScholarship (California Digital Library) · 2014
RedCooper: Hardware Sensor Enabled Variability Software Testbed for Lifetime Energy Constrained Application Yuvraj Agarwal † , Alex Bishop † , Tuck-Boon Chan † , Matt Fotjik ‡ , Puneet Gupta ∗ , Andrew B. Kahng † , Liangzhen Lai ∗ , Paul Martin ∗ , Mani Srivastava ∗ , Dennis Sylvester ‡ , Lucas Wanner ∗ and Bing Zhang ∗ UCLA ∗ , UCSD † , University of Michigan ‡ E-mails: [email protected] Abstract—Conventional hardware uses overdesigned mar- gins to guardband against variability, which incurs significant amounts of power and performance overhead. If the varia- tions can be captured and exposed to the higher levels (e.g., system/software levels), the margin can be reduced or even eliminated through opportunistic hardware/software adaptation. In this paper, we present our end-to-end implementation of an software testbed with built-in hardware sensors and adaptive software. The measurement results show that using our novel performance sensor, Design-Dependent Ring-Oscillator (DDRO), can reduce the mean delay estimation errors by up to 35% (from 2.5% to 1.5%) compared to using generic inverter-based ring-oscillator. By utilizing the sensing infrastructure on our RedCooper testbed, a demonstration shows that the hardware and software adaptation can achieve up to 2.7X total active time increase for lifetime energy constrained, as compared to sensorless system. I. I NTRODUCTION As semiconductor manufacturing process advances and fea- ture size shrinks, hardware sees an increasing amount of variability. Performance variation and power variation are the two major manifestations of hardware variability. Conventional approaches use overdesign and guardbands for the variations, which leaves increasing amounts of power and performance potential untapped. If the variations can be captured and exposed to the higher levels of stacks (e.g., system and software levels), the margin can be reduced or even eliminated through opportunistic hardware/software adaptation [1]. For example, a processor typically has some margin on its operating frequency to guarantee correct functionality across the worst-case process, voltage and temperature (PVT) variations. The margin can be reduced if we can estimate the processor’s performance through on-chip performance monitors (e.g., ring-oscillators). When the performance monitor reports an estimation better than worst-case, which is likely, we can operate the processor at a higher frequency. Alternatively, we can utilize the “better than worst-case” scenario by lowering the supply voltage for power reduction. In addition to hardware adaptations enabled by perfor- mance sensors, software adaptation can also help eliminate the overdesigned guardband. For example, [2] proposes software duty-cycling for lifetime energy constrained embedded sensing applications. By exposing system’s current power consumption information to the software, the application can dynamically adjust its duty cycle rates (i.e., ratio of chips active time to its lifetime) so that overall quality of service (QoS), as determined by system active time, can be optimized within pre-specified energy budget and lifetime requirements. There are three major issues in the implementation of such an adaptive system: 1).How to design efficient and accurate sensors to capture the hardware variability signature? 2). How to expose the hardware sensors to the system and software? 3). How can the system and software adapt to the sensed hardware variability? In this work, we present our end-to-end implementation of such systems which addresses the above implementation issues by: We implement a testchip of Design-Dependent Ring- Oscillator (DDRO), a systematic way of designing and leveraging multiple replica monitors. The silicon mea- surement results show that using multiple DDROs can reduce the mean delay estimation errors by up to 35% (from 2.5% to 1.5%) compared to using generic inverter- based ring-oscillator. We implement a testbed RedCooper based on our testchip, which offers the sensing infrastructure, and more importantly, capability of exposing the sensor readings to the software. We demonstrate an adaptive duty-cycling application based on a embedded operating system running on RedCooper testbed. Our demonstration shows that the hardware and software adaptation can achieve up to 2.7X total active time increase for lifetime energy constrained, as compared to sensorless system. The rest of the paper is organized as follows: Section II describes our implementation of DDRO testchip and present the silicon measurement results. Section III describes the variability-aware software duty-cycling methodology. Sec- tion IV explains our board implementation of the RedCooper testbed. Section V presents our software implementation and the demonstration of application running on the testbed with both hardware and software adaptations. We conclude the paper in Section VI.