Energy consumption of TAGS in a heterogeneous environment under unknown service demand
Introduction
Modern scheduling algorithms require low energy consumption as well as good performance. Having a suitable scheduling algorithm in terms of performance is not enough if it has a negative impact on energy consumption.
Jakóbik et al. [1] discussed several scheduling and job allocation strategies that significantly improve energy efficiency. Lent [2] studied optimal energy consumption in an energy proportional data centre, where servers become available on demand. The primary objective was to analyse the optimal energy requirements of servers when meeting performance service level objectives. Three different scenarios were considered, including servers running at or below maximum utilisation, controlling the average response time setting a specific limit, and reducing incidences of job response times exceeding deadlines.
Ricciardi et al. [3] trialled turning a subset of servers in a data centre off and on to conserve power in an environment of fluctuating service demand. To evaluate the trade-offs when reducing energy consumption.
Phung-Duc [4] considered a multi-server queuing model with setup time and impatient customers to analyse power saving and performance trade-offs at data centres. The server was set to shut down immediately if there was no job, and to switch on when a new job arrives.
Mathew [5] presented an algorithm for cluster shutdowns, that turns off servers in an entire cluster of content delivery networks (CDNs). Real-world traces from a group of commercial CDNs were used to evaluate this technique, and found a reduction of 67% was achievable.
In this paper, we consider the impact of choice of scheduler on performance and energy in a heterogeneous environment. We evaluate task assignment based on guessing size algorithm, known as (TAGS). In [6], [7], the authors focused mostly on the performance of TAGS with respect to mean slowdown target function and mean (normalised) waiting time. In this paper, we examine throughput as a performance metric and study the energy consumption of TAGS. To the best of our knowledge, TAGS energy consumption has not been analysed previously. In [8] we studied the performance and energy consumption of TAGS in a homogeneous environment.
The remainder of this paper is organised as follows: Section 2 illustrates the concept of TAGS. Section 3 presents the TAGS model using the Markovian process algebra PEPA. Section 4 introduces the energy model. Section 5 presents the performance analysis for TAGS in a heterogeneous environment. The energy results are discussed in Section 6.
Section snippets
TAGS algorithm
The TAGS algorithm was initially introduced by Harchol-Balter [6]. It functions principally to allocate jobs where service demand is variable and unknown. Following this approach, a job is first sent to a single server queue. The server starts serving the first job in the queue, continuing until the job is completed and departs successfully, or until a fixed time out is reached. If the time out was reached prior to completion of the job, then the job is immediately transferred to the next queue
TAGS PEPA model in heterogeneous environment
Previous works [9], [8], modelled TAGS in a homogeneous environment, so the PEPA model represents interactions in such an environment. However, considering heterogeneity in the system environment is vital when modelling server's performance and energy consumption. In this case, the PEPA model needs to be redesigned to adapt to changes in the system environment.
The number of nodes is restricted to just two, which is sufficient to investigate the consequences of using this mechanism on energy
The energy model
The choice of the energy model is an integral feature of the study. We focus on a case where the majority of the server's power consumption is by the CPU, and so neglect components such as the memory and hard disk. The process we propose to estimate energy consumed Ec by a server assumes energy is essentially processor performance states (P-states) multiplied by utilisation. To estimate energy per job, the throughput of the system is combined with the utilisation and P-state value, to obtain an
TAGS performance analysis in heterogeneous environment
This section illustrates our ongoing work reviewing the performance of the TAGS algorithm under hyper-exponential service demand in a heterogeneous environment.
Energy consumption analysis
This section illustrates our ongoing work evaluating energy consumption with the TAGS algorithm, the shortest queue and the weighted random. We studied three algorithms’ energy consumption under hyper-exponential service demand in a heterogeneous environment.
Conclusion
This paper studied the performance and energy consumption of the TAGS algorithm in a heterogeneous environment, comparing energy results with shortest queue and weighted random.
We focused on CPU energy consumption and neglected other server components. We used P-state values to ascertain energy consumption. We considered three combinations of servers: identical; first server slower than the second server; and second server slower than the first server.
The analysis reveals that the TAGS
Authors’ contribution
Ali Alssaiari: conceptualisation, methodology, investigation, writing-reviewing and editing. Nigel Thomas: supervision, methodology, writing-reviewing and editing.
Conflict of interest
None declared.
Declaration of Competing Interest
The authors report no declarations of interest.
References (11)
Analysis of an energy proportional data center
Ad Hoc Netw.
(2015)- et al.
Analysis of the task assignment based on guessing size policy
Perform. Eval.
(2020) Comparing job allocation schemes where service demand is unknown
J. Comput. Syst. Sci.
(2008)- et al.
Energy efficient scheduling methods for computational grids and clouds
J. Telecommun. Inform. Technol.
(2017) - et al.
Saving energy in data center infrastructures