首页|Working Set Size Estimation Techniques in Virtualized Environments: One Size Does not Fit All
Working Set Size Estimation Techniques in Virtualized Environments: One Size Does not Fit All
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
Energy consumption is a primary concern for datacenter (DC) management. Its cost represents a significant part of the total cost of ownership (about 80% [2]) and it is estimated that in 2020, US DCs will spend about $13 billion on energy bills [3].A majority of DCs implements the Infrastructure as a Service (IaaS) model where customers buy (from providers) VMs with a set of reserved resources. The VMs host general purpose applications (e.g. web services), as well as High Performance Computing applications. In such IaaS DCs, virtualization is a fundamental technology which allows optimizing the infrastructure by colocating several VMs on the same physical server. Such colocation can be achieved at deployment time by starting as many VMs as possible on each physical machine, or at runtime by dynamically migrating VMs on a reduced set of physical machines, thus implementing a consolidation strategy [4].Ideally, consolidation should lead to highly loaded servers. Although consolidation may increase server utilization by about 5-10%, it is difficult to actually observe server loads greater than 50% for even the most adapted workloads [5-7]. The main reason is that VM collocation is memory bound, as memory saturates much faster than the CPU. This situation was accentuated over the last several years, as we have seen emerging new applications with growing memory demands, while physical platforms had an opposite tendency; they provide more CPU capacity than physical memory. This mismatch is referred to as the memory capacity wall [8].However, the existing consolidation systems [9,10] take the CPU as a pivot, i.e. the central element of the consolidation. The memory is considered constant (i.e. the initially booked value) all over the VM's lifetime. Nevertheless, we consider that the memory should be the consolidation pivot since it is the limiting resource. In order toreduce the memory pressure, the consolidation should consider the memory actually consumed (i.e. the VMs working set size) and not the booked memory (see Fig. 1). Thereby, we need mechanisms to (1) evaluate the working set size (WSS) of VMs, (2) to anticipate their memory evolution and (3) to dynamically adjust the VMs' allocated memory. Numerous research papers propose algorithms to estimate the WSS of VMs. However, most of them are able to follow either up-trends (the increase) or down-trends (the decrease) of WSS. The few of them which are able to follow both trends are highly intrusive. Moreover, to the best of our knowledge, no previous work has shown the implications of dynamically adjusting the VMs allocated memory according to the WSS estimation. Finally, as far as we know, no previous consolidation algorithm considers the WSS as a pivot. In this paper we address all the above limitations. In summary, the contributions of this paper are the following:1.We define evaluation metrics that allow to characterize WSS estimation solutions.2.We evaluate existing WSS techniques on several types of benchmarks. Each solution was implemented in the Xen virtualization system.3.We propose Badis, a WSS monitoring and estimation system which leverages several of the existing solutions in order to provide high estimation accuracy with no codebase intru-siveness. Badis is also able to dynamically adjust the VMs allocated memory based on the WSS estimations.4.We propose a consolidation system extension which leverages Badis for a better consolidation ratio. Both the source and the data sets used for our evaluation are publicly available [1], so that our experiments can be reproduced.
Toulouse University Toulouse Institute of Computer Science Research Toulouse, France,Toulouse University Toulouse Institute of Computer Science Research Toulouse, France
Toulouse University Toulouse Institute of Computer Science Research Toulouse, France
Institute for Informatics and Automation Problem Yerevan, Armenia
NETL
NSTL
