A Study on Acceleration Methods of Data Center Applications with Reconfigurable Hardware
福田 駿
2014 年度 卒 ／博士（工学）
平成26年度 日本学術振興会特別研究員
博士論文の概要
This study discusses the use of reconfigurable hardware, especially in data centers.
Reconfigurable hardware is a promising technology for overcoming the difficulties face by general-purpose processors, which have been central to computer science for decades. Enabling reconfigurable technology to be used by a variety of people, including software developers, will have a great impact on pushing current computing to a new era. The recent enthusiasm for cloud computing indicates the increasing significance of data centers, and improving the performance of data centers in computation speed and energy efficiency has a great impact on the entire computer industry.
Although computers have progressed enormously since their invention, physical restriction is beginning to prevent them from achieving higher computation speed and energy efficiency. These two matrices, computation speed and energy efficiency, are critical in data centers that handle gigantic data traffic from all over the world. As the amount of data that data centers handle is growing even more, computer developers need different ways of improving computers. Among the several alternatives that have been proposed, reconfigurable hardware is one of the most suitable choices, and this is what we try to use in this study. Reconfigurable hardware promises higher performance and energy efficiency in many application domains, and several research groups have been trying to deploy it in data centers.
Meanwhile, reconfigurable hardware is used mainly by embedded or networking device vendors, and curious end users often simply enjoy trying them with reconfigurable boards purchased on their own. This is because it is difficult to obtain higher performance with inexpensive reconfigurable boards, which an individual can buy, than with a commercially available general-purpose processor, and reconfigurable boards that can exceed the performance of commercial general-purpose processor are too expensive for end users to buy individually. By deploying reconfigurable hardware and making it available to the public, the cost of utilizing reconfigurable hardware will go down and thus more end-users will be eager to use it. More applications such as databases that require very low latency will be built on reconfigurable hardware as well as general-purpose processors. Such a computing system, which uses several general-purpose processors and reconfigurable devices or heterogeneous systems, is too complicated for end users' everyday use.
As a consequence, to make the system look simple for the end users, it will become more transparent: they will eventually not notice which kind of computing device they are using. However, there are many problems that must be solved before such a cloud system can be widely used. Among them, we focus on two major problems in this study: developing applications with reconfigurable hardware is more difficult compared to software applications, thus, the cost of developing the application is higher; and a method for using reconfigurable hardware in data centers is not established.
In order to solve the first problem, we assess where the difficulties lie in the state of the art design method that uses high-level synthesis tools for developing hardware accelerated application systems, and propose a method to overcome the difficulties. For solving the other problem, we propose a new technique of using reconfigurable hardware in data centers.