CARB: A C-State Power Management Arbiter for Latency-Critical Workloads

Zhan, Xin; Azimi, Reza; Kanev, Svilen; Brooks, David; Reda, Sherief

doi:10.1109/lca.2016.2537802

Cited by 18 publications

(11 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…One common host-level power saving technique, DVFS, can lower the power consumption by running workloads more slowly. DVFS policies and mechanisms have become more sophisticated over time, starting from power throttling of a chip multi-processor [21], to finer-grain per-core management [37,39], to workload-and request-aware throttling decisions [9,10,20,23,31,45,46]. Newer approaches that can work in high-utilization environments generally require visibility into application-exported metrics.…”

Section: Why Consider Applications As Opaque Boxes?mentioning

confidence: 99%

Leveraging application classes to save power in highly-utilized data centers

Kaffes

Sbîrlea

Lin

et al. 2020

Proceedings of the 11th ACM Symposium on Cloud Computing

View full text Add to dashboard Cite

Data center energy consumption has become an increasingly significant contributor both to greenhouse emissions and costs. To increase utilization of individual hosts and improve efficiency, most modern data centers co-locate workloads belonging to different application classes, some being latency-sensitive (LS) and others best-effort (BE) which are more tolerant to performance variation. It is therefore necessary to design mechanisms that reduce power consumption even in the resulting high-utilization environment, while preserving LS task performance. Moreover, the abundance of different workloads and the security implications of public cloud make mechanisms that rely on extensive knowledge of workload characteristics or on application-exported metrics challenging to deploy. We present PACT, Per Application Class Turbo Controller, a system that leverages two novel mechanisms to reduce power consumption even in highly-utilized data centers. By treating applications like opaque boxes that do not need to provide application-specific performance signals, the first mechanism, Turbo Control, reduces power consumption by decreasing the operating frequency and throttling only BE tasks, without affecting performance-sensitive LS tasks. We identify the shortcomings of Turbo Control and increase its effectiveness by introducing CPUJailing, a mechanism that allocates different sets of cores to LS and BE applications. We deploy PACT (Turbo Control + CPUJailing) in production * Kostis Kaffes was an intern at Google during this work. † Christos Kozyrakis was partly at Google during this work.

show abstract

Section: Why Consider Applications As Opaque Boxes?mentioning

confidence: 99%

Leveraging application classes to save power in highly-utilized data centers

Kaffes

Sbîrlea

Lin

et al. 2020

Proceedings of the 11th ACM Symposium on Cloud Computing

View full text Add to dashboard Cite

show abstract

“…Feedback-based controllers. Feedback-based controllers periodically measure the incoming rate or the observed latency, compare them to predefined thresholds, and adjust the number of cores allocated to the KV store application [36,64].…”

Section: Existing Systemsmentioning

confidence: 99%

“…Rubik adjusts frequency upon arrival and completion of each request. Their simulation results indicate that Rubik has better power consumption and tail latency than approaches relying on feedback-based controllers (e.g., Carb [64] and Pegasus [36]) since Rubik is responsive enough to cope with short term variations. • DPM [8] extends idle period lengths by consolidating idle periods through delayed request processing.…”

Section: Evaluated Systemsmentioning

confidence: 99%

“…Feedback based approaches. Pegasus [36] and Carb [64] leverage feedback-based controllers for saving power. Pegasus periodically measures tail latency and adjusts the frequency every few seconds to avoid tail latency violations.…”

Section: Related Workmentioning

confidence: 99%

“…One approach to building energy-proportional KV stores is to use feedback-based controllers that monitor and measure the load, compare it to predefined thresholds, and scale the number of active cores to save power during periods of low or medium utilization [36,64]. Feedback-based controllers can adapt to diurnal variations; however, they are too slow (operating at second-scale intervals) to cope with short-term burstiness, lengthening tail latency [27].…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Peafowl

Asyabi

Bestavros

Sharafzadeh

et al. 2020

Proceedings of the 11th ACM Symposium on Cloud Computing

View full text Add to dashboard Cite

The traffic load sent to key-value (KV) stores varies over long timescales of hours to short timescales of a few microseconds. Long-term variations present the opportunity to save power during low or medium periods of utilization. Several techniques exist to save power in servers, including feedback-based controllers that right-size the number of allocated CPU cores, dynamic voltage and frequency scaling (DVFS), and c-state (idle-state) mechanisms. In this paper, we demonstrate that existing power saving techniques are not effective for KV stores. This is because the high rate of traffic even under low load prevents the system from entering low power states for extended periods of time. To achieve power savings, we must unbalance the load among the CPU cores so that some of them can enter low power states during periods of low load. We accomplish this by introducing the notion of in-application CPU scheduling. Instead of relying on the kernel to schedule threads, we pin threads to bypass the kernel CPU scheduler and then perform the scheduling within the KV store application. Our design, Peafowl, is a KV store that features an in-application CPU scheduler that monitors the load to learn the workload characteristics and then scales the number of active CPU cores when the load drops, leading to notable power savings during low or medium periods of utilization. Our experiments demonstrate that Peafowl uses up to 40-54% lower power than state of the art approaches such as Rubik and DPM.

show abstract

Co-Adjusting Voltage/Frequency State and Interrupt Rate for Improving Energy-Efficiency of Latency-Critical Applications

Kang

Park

et al. 2020

IEEE Access

View full text Add to dashboard Cite

As the power/energy consumption is one of the major contributors to the Total Cost of Ownership (TCO), improving power/energy efficiency is crucial for large-scale data centers where latencycritical applications are commonly accommodated while computing resources are usually under-utilized. For improving the power/energy efficiency of processors, most of the commercial processors support Dynamic Voltage and Frequency Scaling (DVFS) technology that enables to adjust Voltage and Frequency state (V/F state) of the processor dynamically. In particular, for the latency-critical applications, many prior studies propose power management policies using the DVFS for the latency-critical applications, which minimizes the performance degradation or satisfies the Service Level Objectives (SLOs) constraints. Meanwhile, although the interrupt rate also affects the response latency and energy efficiency of latencycritical applications considerably, those prior studies just introduce policies for V/F state adjustment while not considering the interrupt rate. Therefore, in this paper, we investigate the impact of adjusting the interrupt rate on the tail response latency and energy consumption. Through our experimental results, we observe that adjusting interrupt rate along with V/F state management varies the performance and energy consumption considerably, and provides an opportunity to reduce energy further by showing latency overlap between different V/F states. Based on the observation, we show the quantitative potential in improving energy efficiency of co-adjusting V/F state and interrupt rate with a simple management policy, called Co-PI. Co-PI searches the most energy-efficient combination of the V/F state and interrupt rate from the latency and energy tables that we obtain through offline profiling, and reflect the combination to the core and NIC. Co-PI reduces energy consumption by 34.1% and 25.1% compared with performance and ondemand governors while showing the almost same tail response latency with the performance governor that operates cores at the highest V/F state statically.

show abstract

CARB: A C-State Power Management Arbiter for Latency-Critical Workloads

Cited by 18 publications

References 7 publications

Leveraging application classes to save power in highly-utilized data centers

Leveraging application classes to save power in highly-utilized data centers

Peafowl

Co-Adjusting Voltage/Frequency State and Interrupt Rate for Improving Energy-Efficiency of Latency-Critical Applications

Contact Info

Product

Resources

About