Selective switching mechanism in virtual machines via support vector machines and transfer learning

Abstract

Virtualization is an essential technology in data centers allowing for a single machine to be used for multiple applications or users. With memory virtualization, two approaches, shadow paging (SP) and hardware-assisted paging (HAP), are taken by modern virtual machine memory managers. Neither memory mode is always preferred; previous studies have proposed to exploit the advantages of both modes by dynamically switching between these two paging modes based on the on-the-fly system behavior. However, the existing scheme makes the switching decision based on manual rules summarized for a specific architecture. This paper employs a machine learning approach that learns a decision model automatically and thus can adapt to different systems. Experimental results show that the performance of our switching mechanism can match or outperform either SP or HAP alone. Also, the results demonstrate that a machine learning-based decision model can match the performance of the hand-tuned model. Moreover, we further show that different hardware/software settings can affect on-the-fly system behavior and thus demand different decision models. Our scheme yields two effective decision models on two different machines. Additionally, transfer learning was used in order to efficiently train a model when faced with a new hardware configuration with only a limited number of training samples from the new machine.

Constructing Dynamic Policies for Paging Mode Selection

“…This section describes our experimental methodology and presents our results. We evaluate the performance of both DSP-SAMPLE and DSP-OFFSET, and compare both models against the state-ofthe-art ASP-SVM [16]. To conclude, we discuss the advantages of DSP-OFFSET with respect to profiling cost (Section 4.4).…”

“…Kuang et al [16] designed a procedure for labeling program phases according the performance gain associated with each paging mode, and utilize machine learning to construct a decision procedure. They enumerate over each phase of a program, comparing the performance of selecting HAP for that phase (and SP for the remaining phases) with the baseline performance of SP; similarly, they enumerate and compare SP with the baseline performance of HAP.…”

“…As with previous work [1,16,31], we rely on page faults and DTLB miss counts to characterize the relative performance of HAP and SP. The frequency of DTLB misses is correlated with the frequency of page walks, and the frequency of page faults is correlated with page table updates; therefore, we expect HAP to outperform SP during periods of frequent page faults and SP to outperform HAP during periods of frequent DTLB misses.…”

“…Previous work has proposed dynamic methods for selecting between HAP and SP at runtime depending on workload performance characteristics, using manual analysis and a hand-tuned model [31] or expensive enumerative profiling and machine learning [16]. Both cases show that dynamic selection can improve performance by matching, and in some cases beating, the performance of a static paging choice.…”

“…We map the problem of selective paging to the contextual bandit, a model for sequential decision making under limited feedback. With a single, random profiling execution of each benchmark in the SPEC INT2006 suite, using the Binary-Offset algorithm [5], we construct an effective dynamic paging mode selection policy which is competitive with the state-of-the-art ASP-SVM [16] while requiring substantially less profiling time. Unlike previous work, our profiling requires no prior knowledge of workload structure or phasing, and does not require extensive domain expertise or manual tuning.…”