Intf-HybridMem: Page migration in hybrid memories considering cost efficiency

Highlights

• This paper presents the Intf-HybridMem architecture, a mechanism for managing page migration in hybrid main memory using a fuzzy approach.

• The main objective of this work is to evaluate the potential of the Intf-HybridMem architecture in its efficiency considering the hybrid memories characteristics.

• Tests based on the oracle migration policy show that there is potential to better explore the hybrid memories and its features of low energy consumption.

Abstract

One of the great challenges of today's computer memory architectures has been their energy consumption. In this sense, DRAM technology has come close to its scalability limit due to energy issues. Thus, new technologies have emerged intending to replace DRAM as main memory. In this context, non-volatile memories emerge promising to be an alternative for low energy consumption memory systems. Due to the challenges of these new technologies, hybrid approaches have been a trend. When working with hybrid memories, there is a necessity for decisions in which memory modules each data should be stored. This decision must consider the data access profile and the characteristics of memory technology. Thus, this work presents the Intf-HybridMem architecture, a proposal for page migration in hybrid memories using fuzzy systems to support decision making. The fuzzy approach can model the uncertainties of the data access profile and the characteristics of the memory modules. The tests seek to identify the performance limit of the Intf-HybridMem through the use of an oracle mechanism. Results show that there is potential for exploring migration mechanisms with overhead of 0.18% and 0.39% in some scenarios.

Introduction

Main memory capacity has become a limitation in the development of new computational systems. Cloud computing, high-scale computing systems, and big data processing are examples of the challenges that are being assigned to memory architectures. Also, there is a growing need for the development of sustainable computer systems, committed to low energy consumption. Relevant studies in the literature show that DRAM memory technology is achieving its limit of scalability [1], [2]. The main constraints of DRAM technologies are their static energy consumption and the difficulty of reducing the size of memory cells. Moreover, the charge time of capacitors increases as the size of the technology decreases [3], [4], [5]. Consequently, in DRAM technology, the grown of main memory to meet large scale computing leads to an increase in energy consumption and also in power dissipation.

Considering this scenario, new technologies have been developed as alternatives to the DRAM architecture in order to achieve high memory capacity demand. Thus, a strong trend in memory architecture is the non-volatile memories (NVMs), providing lower static power and higher density when compared to DRAM. Besides that, in some cases, NVMs can achieve an overall low energy consumption. Because of these main characteristics, NVM technologies have the potential to overcome the scalability limits related to DRAM memories.

NVM application over traditional memory architectures leads to new challenges because of the low endurance of NVM and the asymmetry time executions between read and write operations. Moreover, in these scenarios, where NVM is still under development and the simple replacement of DRAM by NVM architecture would face durability issues [6], the use of hybrid memories constitutes a new perspective to prospect solutions of such memory architecture problems. This proposal integrates the use of both approaches, DRAM and NVM, exploiting the best features of each technology [7].

One basic problem in the hybrid memory research area is the following decision-question: Which does memory module need to be select for each data writing in a request service? This decision-making problem should has to consider the uncertainties inherent to multiple characteristics of hybrid memory architectures. See, e.g., the size of memories, speed of read/write operations, endurance, and energy consumption. Moreover, and more challenging, the decision should also has to examine and identify the patterns of memory accesses, also considering the uncertainty related to all previously described attributes.

The conception of Intf-HybridMem architecture and its evaluation is presented as a proposal for migrating pages in hybrid main memory using a flexible approach based on interval-valued fuzzy logic. Fuzzy systems seek to model the uncertainties inherent of hybrid memories considering both aspects, main characteristics and distinct behavior of memory accesses, in order to support decision making problems mitigating energy consumption. And, for analyzing the performance of limit evaluations related to the Intf-HybridMem architecture, the use of an oracle migration policy is considered. Tests based on the oracle migration policy show that there is potential to better explore the main features of hybrid memories in order to achieve low energy consumption.

Section 1 shows an overview of memory technologies and hybrid memory approaches. Section 3 discuss related works. Fuzzy theory and foundations which support this work are presented in Section 2.2. Section 4 presents the architectural organization of the Intf-HybridMem. Tests, evaluation and results are discussed on Section 5, and conclusion and future works in Section 6.