Modelling Energy Efficiency and Performance Trade-offs

Abstract

Power and energy consumption in data centres is a huge concern for data centre providers. As a result, this work considers the modelling and analysis of policy and scheduling schemes using Markovian processing algebra known as PEPA. The first emphasis was on modelling an energy policy in PEPA that dynamically controls the powering servers ON or OFF. The focus is to identify and reflect the trade-off between saving energy (by powering down servers) and performance cost. While powering down servers saves energy, it could increase the performance cost. The research analyses the effect of the policy on energy consumption and performance cost, with different combinations of dynamic and static servers used in the policy against different scenarios, including changes in job arrival rate, job arrival duration and the time needed by servers to be powered On and start process jobs. The result gave interesting outcomes because every scenario is unique, and therefore, no server combinations were found to give low energy and high performance in all situations.

The second focus was to consider the impact of scheduler’s choice on performance and energy under unknown service demands. Three algorithms were looked at: task assignment based on guessing size (TAGS), the shortest queue strategy and random allocation. These policies were modelled using PEPA to derive numerical solutions in a two servers system. The performance was analysed considering throughput, average response time and servers' utilisation. At the same time, the energy consumption was in terms of total energy consumption and energy consumption per job. The intention was to analyse the performance and energy consumption in a homogeneous and heterogeneous environment, and the environment was assumed to be homogeneous in the beginning. However, the service distribution was considered either a negative exponential (hence relatively low variance) or a two-phase hyper-exponential (relatively high variance) in each policy. In all cases, the arrival process has been assumed to be a Poisson stream, and the maximum queue lengths are finite (maximum size is 10 jobs). The performance results showed that TAGS performs worse under exponential distribution and the best under two-phase hyper-exponential. TAGS produce higher throughput and lower job loss when service demand has an H2 distribution. Our results show that servers running under TAGS consume more energy than other policies regarding total energy consumption and energy per job under exponential distribution. In contrast, TAGS consumes less energy per job than the random allocation when the arrival rate is high, and the job size is variable (two-phase hyper-exponential).

In a heterogeneous environment and based on our results on the homogeneous environment, the performance metrics and energy consumption was analysed only under two phase hyper-exponential. TAGS works well in all server configurations and achieves greater throughput than the shortest queue or weighted random, even when the second Server's speed was reduced by 40% of the first server's in TAGS. TAGS outperforms both the shortest queue and weighted random, whether their second server is faster or slower than the TAGS second server. The system's heterogeneity did not significantly improve or decrease TAGS throughput results. Whether the second server is faster or slower, even when the arrival rate is less than 75% of the system capacity, it approximately showed no effect. On the other hand, heterogeneity of the system has a notable effect on the throughput of the shortest queue and weighted random. The decrease or increase in throughput follows the trend of the second server performance capability. In terms of total energy consumption, for all scheduling schemes, when the second server is slower than the first server, the energy consumption is the highest among al