Analysis of power‐saving techniques over a large multi‐use cluster with variable workload

Abstract

SUMMARYReduction of power consumption for any computer system is now an important issue, although this should be done in a manner that is not detrimental to the users of that computer system. We present a number of policies that can be applied to multi-use clusters where computers are shared between interactive users and high-throughput computing. We evaluate policies by trace-driven simulations in order to determine the effects on power consumed by the high-throughput workload and impact on high-throughput users. We further evaluate these policies for higher workloads by synthetically generating workloads based around the profiled workload observed through our system. We demonstrate that these policies could save~45% of the currently used energy for our high-throughput jobs over our current cluster policies without affecting the high-throughput users experience.

HTC‐Sim: a trace‐driven simulation framework for energy consumption in high‐throughput computing systems

“…The HTC system must select the most appropriate candidates from a selection of avialable resources, in order to optimise the required performance, energy and QoS metrics. Many approaches exist including random allocation, reduced energy consumption, reduced likelihood of eviction, or fastest execution time [56,57,58] -discussed further in Section 5.5. The Pluggable Policy interface provides a simple interface which a developer can use to implement their own scheduler.…”

“…Cluster Policy in Figure 4) The infrastructure on which many Desktop Grids are installed have nightly reboot policies allowing for system clean-up and software updates. Given that HTC workloads tend to have the greatest chance of running to completion throughout the night, particularly true for longer-running jobs, the ability for HTC jobs to defer these reboots can significantly improve the chance of jobs completing [56,58].…”

“…Selecting an appropriate suspension timeout, after which the HTC job is abandoned and commences execution on a different computing node, represents a challenging trade-off. If the suspension timeout is too short then this benefit can be lost whilst if the timeout is too long then a significant penalty is imposed on the time a job takes to complete [56,58].…”

“…This policy was first introduced in [70] as an extension of resource allocation strategies first explored in [56].…”

“…In order to evaluate the scalability of our simulation software we investigate the execution time and memory footprint when running larger (synthetic) workloads [56] -over ten times our real Figures 17 and 18 show the memory footprint and execution times respectively for both our original simulation and synthetic trace logs under policies S1 and S4. Error bars show the standard deviation between based on fifteen simulation runs.…”