An official website of the United States government
Power capping is an important technique for high-density servers to safely oversubscribe the power infrastructure in a data center. However, power capping is commonly accomplished by dynamically lowering the server processors’ frequency levels, which can result in degraded application performance. For servers that run important machine learning (ML) applications with Service-Level Objective (SLO) requirements, inference performance such as recognition accuracy must be optimized within a certain latency constraint, which demands high server performance. In order to achieve the best inference accuracy under the desired latency and server power constraints, this paper proposes OptimML, a multi-input-multi-output (MIMO) control framework that jointly controls both inference latency and server power consumption, by flexibly adjusting the machine learning model size (and so its required computing resources) when server frequency needs to be lowered for power capping. Our results on a hardware testbed with widely adopted ML framework (including PyTorch, TensorFlow, and MXNet) show that OptimML achieves higher inference accuracy compared with several well-designed baselines, while respecting both latency and power constraints. Furthermore, an adaptive control scheme with online model switching and estimation is designed to achieve analytic assurance of control accuracy and system stability, even in the face of significant workload/hardware variations.
more »
« less
PADS: Power Budgeting with Diagonal Scaling for Performance-Aware Cloud Workloads
Cloud platforms’ rapid growth raises significant concerns about their electricity consumption and resulting carbon emissions. Power capping is a known technique for limiting the power consumption of data centers where workloads are hosted. Today’s data center computer clusters co-locate latency-sensitive web and throughput-oriented batch workloads. When power capping is necessary, throttling only the batch tasks without restricting latency-sensitive web workloads is ideal because guaranteeing low response time for latency-sensitive workloads is a must due to Service-Level Objectives (SLOs) requirements. This paper proposes PADS, a hardware-agnostic workload-aware power capping system. Due to not relying on any hardware mechanism such as RAPL and DVFS, it can keep the power consumption of clusters equipped with heterogeneous architectures such as x86 and ARM below the enforced power limit while minimizing the impact on latency-sensitive tasks. It uses an application-performance model of both latency-sensitive and batch workloads to ensure power safety with controllable performance. Our power capping technique uses diagonal scaling and relies on using the control group feature of the Linux kernel. Our results indicate that PADS is highly effective in reducing power while respecting the tail latency requirement of the latency-sensitive workload. Furthermore, compared to state-of-the-art solutions, PADS demonstrates lower P95 latency, accompanied by a 90% higher effectiveness in respecting power limits.
more »
« less
- PAR ID:
- 10591371
- Publisher / Repository:
- IEEE
- Date Published:
- ISBN:
- 979-8-3315-0786-2
- Page Range / eLocation ID:
- 14 to 21
- Format(s):
- Medium: X
- Location:
- Austin, TX, USA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
null (Ed.)With the rapid development of the Internet of Things (IoT), computational workloads are gradually moving toward the internet edge for low latency. Due to significant workload fluctuations, edge data centers built in distributed locations suffer from resource underutilization and requires capacity underprovisioning to avoid wasting capital investment. The workload fluctuations, however, also make edge data centers more suitable for battery-assisted power management to counter the performance impact due to underprovisioning. In particular, the workload fluctuations allow the battery to be frequently recharged and made available for temporary capacity boosts. But, using batteries can overload the data center cooling system which is designed with a matching capacity of the power system. In this paper, we design a novel power management solution, DeepPM, that exploits the UPS battery and cold air inside the edge data center as energy storage to boost the performance. DeepPM uses deep reinforcement learning (DRL) to learn the data center thermal behavior online in a model-free manner and uses it on-the-fly to determine power allocation for optimum latency performance without overheating the data center. Our evaluation shows that DeepPM can improve latency performance by more than 50% compared to a power capping baseline while the server inlet temperature remains within safe operating limits (e.g., 32°C).more » « less
-
Since emerging edge applications such as Internet of Things (IoT) analytics and augmented reality have tight latency constraints, hardware AI accelerators have been recently proposed to speed up deep neural network (DNN) inference run by these applications. Resource-constrained edge servers and accelerators tend to be multiplexed across multiple IoT applications, introducing the potential for performance interference between latency-sensitive workloads. In this article, we design analytic models to capture the performance of DNN inference workloads on shared edge accelerators, such as GPU and edgeTPU, under different multiplexing and concurrency behaviors. After validating our models using extensive experiments, we use them to design various cluster resource management algorithms to intelligently manage multiple applications on edge accelerators while respecting their latency constraints. We implement a prototype of our system in Kubernetes and show that our system can host 2.3× more DNN applications in heterogeneous multi-tenant edge clusters with no latency violations when compared to traditional knapsack hosting algorithms.more » « less
-
Spiking Neural Networks (SNNs) are an emerging computation model that uses event-driven activation and bio-inspired learning algorithms. SNN-based machine learning programs are typically executed on tile-based neuromorphic hardware platforms, where each tile consists of a computation unit called a crossbar, which maps neurons and synapses of the program. However, synthesizing such programs on an off-the-shelf neuromorphic hardware is challenging. This is because of the inherent resource and latency limitations of the hardware, which impact both model performance, e.g., accuracy, and hardware performance, e.g., throughput. We propose DFSynthesizer, an end-to-end framework for synthesizing SNN-based machine learning programs to neuromorphic hardware. The proposed framework works in four steps. First, it analyzes a machine learning program and generates SNN workload using representative data. Second, it partitions the SNN workload and generates clusters that fit on crossbars of the target neuromorphic hardware. Third, it exploits the rich semantics of the Synchronous Dataflow Graph (SDFG) to represent a clustered SNN program, allowing for performance analysis in terms of key hardware constraints such as number of crossbars, dimension of each crossbar, buffer space on tiles, and tile communication bandwidth. Finally, it uses a novel scheduling algorithm to execute clusters on crossbars of the hardware, guaranteeing hardware performance. We evaluate DFSynthesizer with 10 commonly used machine learning programs. Our results demonstrate that DFSynthesizer provides a much tighter performance guarantee compared to current mapping approaches.more » « less
-
Traditional power reduction techniques such as DVFS or RAPL are challenging to use with web services because they significantly affect the services’ latency and throughput. Previous work sug- gested the use of controllers based on control theory or machine learning to reduce performance degradation under constrained power. However, generating these controllers is challenging as ev- ery web service applications running in a data center requires a power-performance model and a fine-tuned controller. In this paper, we present DDPC, a system for autonomic data-driven controller generation for power-latency management. DDPC automates the process of designing and deploying controllers for dynamic power allocation to manage the power-performance trade-offs for latency- sensitive web applications such as a social network. For each application, DDPC uses system identification techniques to learn an adaptive power-performance model that captures the application’s power-latency trade-offs which is then used to generate and deploy a Proportional-Integral (PI) power controller with gain-scheduling to dynamically manage the power allocation to the server running application using RAPL. We evaluate DDPC with two realistic latency-sensitive web applications under varying load scenarios. Our results show that DDPC is capable of autonomically generating and deploying controllers within a few minutes reducing the ac- tive power allocation of a web-server by more than 50% compared to state-of-the-art techniques while maintaining the latency well below the target of the application.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- https://doi.org/10.1109/IGSC64514.2024.00012
