Tuesday, 12 December 2017

Session 5: Performance and Energy Reduction - Part 2

Energy-Delay Based Provisioning for Large Datacenters: An Energy-Efficient and Cost Optimal Approach

Authors:

Sriram Sankar (Microsoft Corporation)
Kushagra Vaid (Microsoft Corporation)
Harry Rogers (Microsoft Corporation)

Abstract:

It is challenging to determine  the optimum number of servers required to provision for large online applications because of the conflicting mandates of (a) achieving peak performance needs and (b) minimizing unused datacenter power and capacity.  Since Online Services application loads are unpredictable, datacenter operators often conservatively provision for maximum power utilization by characterizing workloads for peak load performance. In contrast, we aim to optimize the service capacity per total cost of ownership (TCO) of an Online Service datacenter deployment by characterizing the energy-delay properties of large scale datacenter workloads. We show that the peak load performance is not the energy efficient point of operation for most applications. We choose two industry-strength workloads (Internet Search and D-Process) and analyze their energy-delay behavior under varying loads. We then calculate the optimal operating point for the specific large-scale application and provision datacenter energy and capacity based on the energydelay curves.  In contrast to workload-based peak power provisioning, we show a 7% benefit in Service Capacity-perTCO-dollar for the energy-delay characterization methodology in our cost analysis for Online Services Applications. 

DOI: 10.1145/1958746.1958778

Full text: PDF

[#][]

Optimizing Benchmark Configurations for Energy Efficiency

Authors:

Meikel Poess Oracle Corporation
Raghunath Nambiar Cisco Systems, Inc.
Kushagra Vaid Microsoft

Abstract:

Historically compute server performance has been the most important pillar in the evaluation of datacenter efficiency, which can be measured using a variety of industry standard benchmarks. With the introduction of industry standard servers, priceperformance became the second pillar in the ‘efficiency equation’. Today with an increased awareness in the industry for power optimized designs and corporate  initiatives to reduce carbon emissions, data center efficiency needs to incorporate yet another key element in this equation: energy efficiency. Initial models based on ‘name-plate’ power consumption have been used to estimate energy efficiency [3][6][8] while recently industry standard consortia like SPEC, TPC and SPC have started amalgating new energy metrics with their traditional performance metrics. TPC-Energy, enables the measuring and reporting of energy efficiency for transaction processing systems and decision support systems [17]. In this paper we analyze TPC-C benchmark configurations that may achieve leadership results in TPC-Energy using existing, more energy efficient technologies, such as solid states drives for storage subsystems, low power processors and high density DRAM in back end server and middle tier systems. Even though the study is based on TPC-C configurations these configuration optimizations are applicable to other benchmarks and production systems alike. We envision that the energy efficiency metrics and related optimizations to claim benchmark leadership will accelerate development and qualifications of energy efficient component and solutions.

DOI: 10.1145/1958746.1958779

Full text: PDF

[#][]

Power and Energy-Aware Processor Scheduling

Authors:

Luigi Brochard (IBM Systems and Technology Group)
Raj Panda (IBM Systems and Technology Group)
Don DeSota (IBM Systems and Technology Group)
Francois Thomas (IBM Systems and Technology Group)
Robert H. Bell, Jr. (IBM Systems and Technology Group)

Abstract:

Power consumption is a critical consideration in high computing systems. We propose a novel job scheduler that optimizes power and energy consumed by clusters when running parallel benchmarks with minimal impact on performance. We construct accurate models for estimating power consumption. These models are based on measurements of power consumption on benchmarks with different characteristics and on systems with processors using different micro-architectures. We show the power estimation models achieve less  than 2% error versus actual measurements.  We show a job  scheduler can be enhanced to make it “power-aware” and to  optimize power consumption of jobs with similar performance characteristics. The enhanced scheduler can estimate the power consumed by a particular job using the power estimation model, configure the nodes in the cluster via suitably adjusting processor frequency on each of the nodes to maximize performance,  minimize power, or minimize energy with a predictable impact on power, energy and performance. 

DOI: 10.1145/1958746.1958780

Full text: PDF

[#][]

Towards More Effective Utilization of Computer Systems

Authors:

Niklas Carlsson (Linköping University)
Martin Arlitt (Hewlett-Packard Laboratories)

Abstract:

Globally, vast infrastructures of Information Technology (IT) equipment are deployed. Much of this infrastructure is under utilized to ensure acceptable response times. This results in less than ideal use of the capital investment used to purchase the IT equipment. To improve the sustainability of IT, we focus on increasing the effective utilization of computer systems. Our results show that computer systems running delay-sensitive (e.g., Web) workloads can be more effectively utilized while still maintaining adequate (e.g., mean or upper percentile) response times. In particular, these computer systems can simultaneously support delay-tolerant workloads, to increase the value of work done by a computer system over time.

DOI: 10.1145/1958746.1958781

Full text: PDF

[#][]