In 2013, U.S. data centers accounted for 2.2% of the country's total electricity consumption, a figure that is projected to increase rapidly over the next decade. Many important workloads are interactive, and they demand strict levels of quality-of-service (QoS) to meet user expectations, making it challenging to reduce power consumption due to increasing performance demands. This paper introduces Hipster, a technique that combines heuristics and reinforcement learning to manage latency-critical workloads. Hipster's goal is to improve resource efficiency in data centers while respecting the QoS of the latency-critical workloads. Hipster achieves its goal by exploring heterogeneous multi-cores and dynamic voltage and frequency scaling (DVFS). To improve data center utilization and make best usage of the available resources, Hipster can dynamically assign remaining cores to batch workloads without violating the QoS constraints for the latency-critical workloads. We perform experiments using a 64-bit ARM big.LITTLE platform, and show that, compared to prior work, Hipster improves the QoS guarantee for Web-Search from 80% to 96%, and for Memcached from 92% to 99%, while reducing the energy consumption by up to 18%.
Given the wide variety of performance demands for various workloads, the trend in embedded systems is shifting from homogeneous to heterogeneous processors, which have been shown to yield performance and energy saving benefits. A typical heterogeneous processor has cores with different performance and power characteristics, that is, high performance and power hungry (“big”) cores, and low power and performance (“small”) cores. In order to satisfy the memory bandwidth and computation demands of various threads, it is important (albeit challenging) to map threads to cores. Such assignment should take into account that threads could potentially be harmful to each other in the usage of shared resources (e.g., cache, memory). We propose a scheme for dynamic energy-efficient assignment of threads to big/small cores, DIO-E (Distributed Intensity Online-Energy), which is an enhancement of the previously proposed DIO. In contrast to DIO, we take into account both CPU and memory demands of threads to characterize the performance of threads when co-running on the same core at run-time. Our results show that DIO-E improves the energy-delay-squared product (ED 2 ) by 9% (average) over DIO, running on a performance-asymmetric multicore system. Both DIO and DIO-E show about 50% improvement in ED 2 over a state-of-the-art solution.
Data centers have time-varying traffic and a wide range of demands in performance-power for the workloads. Understanding the trade-offs between performance and power for these varying types of demands given the time-variations, helps administrators to control the total power consumption and also facilitate for enhancing Quality of Service (QoS) levels. In this paper, we provide a methodology for administrators to predict performance and power using performance monitoring counters at all processor frequency states (P-States) and across a wide range of processor sleep intervals (Cl-States). This methodology will allow administrators to make quick decisions. Specifically, we build models and validate them using SPECcpu2006 benchmarks on an Intel Sandy Bridge processor with 4 cores, 9 P-States and 50 Cl-States resulting in a total of 40 billion configurations when running 4 threads in parallel. The modeling technique provided is fast (under 60,000 cycles), and requires a small period of time to build (less than 10 hours). The modeling technique predicts with a high accuracy and the results obtained show an average error in prediction of 2.25% and 13.99% (with a 95% confidence interval) for power and performance, respectively. Moreover, we provide an insight into the usability of these models in multi-core architectures by predicting performance and power for 3 different types of workloads under 9 different types of QoS constraints. The results obtained show an error in prediction of 9.02% and 7.45% (with a 95% confidence interval) for the total energy and performance delay respectively in multi-core architectures.
Configuration of hardware knobs in multicore environments for meeting performance-power demands constitutes a desirable feature in modern data centers. At the same time, high energy efficiency (performance per watt) requires optimal thread-to-core assignment. In this paper, we present the runtime estimator (RePP-C) for performance-power, characterized by processor frequency states (P-states), a wide range of sleep intervals (Cl-states) and workload consolidation. We also present a schema for frequency and contention-aware thread-to-core assignment (FACTS) which considers various thread demands. The proposed solution (RePP-C) selects a given hardware configuration for each active core to ensure that the performance-power demands are satisfied while using the scheduling schema (FACTS) for mapping threads-to-cores. Our results show that FACTS improves over other state-of-the-art schedulers like Distributed Intensity Online (DIO) and native Linux scheduler by 8.25% and 37.56% in performance, with simultaneous improvement in energy efficiency by 6.2% and 14.17%, respectively. Moreover, we prove the usability of RePP-C by predicting performance and power for 7 different types of workloads and 10 different QoS targets. The results show an average error of 7.55% and 8.96% (with 95% confidence interval) when predicting energy and performance respectively.
One of the main challenges in data center systems is operating under certain Quality of Service (QoS) while minimizing power consumption. Increasingly, data centers are adopting heterogeneous server architectures with different power-performance trade-offs. This requires careful understanding of the application behavior across multiple architectures at runtime so as to enable meeting specified power and performance requirements. In this work, we present and evaluate REPP-H (Runtime Estimation of Performance and Power on Heterogeneous data centers). REPP-H leverages hardware performance counters available on all major server architectures to ensure a highly responsive power capping mechanism and delivering a minimum performance in a single step. We experimentally show that REPP-H can successfully estimate power and performance of several single-threaded and multiprogrammed workloads. The average errors on ARM, AMD and Intel architectures are, respectively, 7.1%, 9.0%, 7.1% when predicting performance, and 6.0%, 6.5%, 8.1% when predicting power on those heterogeneous servers.