Power-Saving Design for Server Farms with Response Time Percentile Guarantees

Abstract

The expense of power cost in server farms has driven the recent power-aware development in both industry and academia. At the same time, a Service Level Agreement (SLA) of service performance between a customer and a service provider is demanded to meet the customer satisfaction. This paper investigates the queueing-theoretical power-saving design strategy for server farms under a given SLA, which in particular is measured in a certain level percentile of the job response time. We consider server farms with servers that are equipped with the capabilities of Dynamic Voltage Scaling (DVFS) and Dynamic Power Management (DPM). We adopt an M/G/1/PS server model, where the job service time distribution is assumed heavytailed, as discovered and validated by previous research. We propose a design strategy called PowerTail to minimize the power consumption under the given SLA. Our data confirms that the proposed PowerTail strategy indeed provides statistical guarantee in comparison with existing dynamic DVFS approaches and significantly outperforms the intuitive load-balancing strategy. 1 Our methodology can be applied to other job scheduling algorithms such as First-Come-First-Served (FCFS) as well.

TailCon: Power-Minimizing Tail Percentile Control of Response Time in Server Clusters

“…1(b). Since the client requests are usually independently generated, we assume that the client requests arrive at the frontend in the server cluster following a Poisson distribution with a mean arrival rate λ F [9], [10]. The frontend dispatches the incoming requests to the backend servers according to a load-balancing policy.…”

“…It is shown that the file sizes of the web requests follow heavy-tailed distributions [13]- [15] and Pareto distribution in particular [9], [15]. We implement a server cluster with 6 Apache servers serving dynamic PhP webpages to the incoming requests, where the client requests are generated following Possion distribution with mean arrival rate λ F and the PhP webpages are generated in a way so that their file sizes X follows Pareto distribution [16],…”

“…However, they did not take the power management into consideration. Wang et al [9] proposed an computing capacity planning strategy to enforce the response time percentile constraint while minimizing power, but it is an offline strategy that cannot be applied for real-time operation of the server cluster. In contrast, we propose TailCon scheme that can dynamically provide the computing capacity online according to the varying workload and can minimize the power consumption subject to the SLA constraint.…”

Delay-Sensitive Service Request Scheduling for Cloud Computing

“…Wang et al [121] proposed a power minimization approach based on the traditional queuing model. Their method determines the appropriate number of powered-on servers and their running modes (i.e.…”

“…The global layer uses bin-packing methods to find a new destination for the virtual machine that has to be migrated. Wang et al [121] propose an approach for power minimization by determining the appropriate numbers of powered-on servers and their running modes (i.e. execution speeds) for the service requests, which can be modeled using the traditional queuing model.…”