著者

Balazs Gerofi Hajime Fujita Yutaka Ishikawa

雑誌

情報処理学会論文誌コンピューティングシステム(ACS) (ISSN:18827829)

巻号頁・発行日

vol.3, no.1, pp.1-12, 2010-03-16

---

Single IP Address cluster offers a transparent view of a cluster of machines as if they were a single computer on the network. In such an environment, process migration can play a significant role for providing services seamlessly and for increasing sustainability. In this paper we propose a live migration mechanism which is capable of moving processes that maintain a massive amount of network connections, supporting both TCP and UDP sockets. Incoming packet loss during socket migration is prevented by exploiting the broadcast property of the Single IP Address cluster, while process live migration minimizes the execution freeze time during the actual migration of the process context. Performance evaluation on machines equipped with a 2.4 GHz CPU and Gigabit Ethernet interconnect shows that migrating a process of 1GB image size and over 1000 established network connections results in less than 200 ms process freeze time, rendering the transition fully transparent and responsive from the clients' point of view. The implementation is comprised entirely of a kernel module for Linux 2.6, without any changes to the existing kernel code.

著者

Masamichi Takagi Balazs Gerofi Norio Yamaguchi Takahiro Ogura Toyohisa Kameyama Atsushi Hori Yutaka Ishikawa

雑誌

研究報告システム・アーキテクチャ(ARC) (ISSN:21888574)

巻号頁・発行日

vol.2015-ARC-215, no.1, pp.1-8, 2015-05-19

---

Processor core count in high-end computing has seen a steady increase during the past decade and next generation supercomputers will likely deploy many-core based systems. At the same time, from a software environment point of view, Linux-compatibility has become wide-spread in the High Performance Computing (HPC) domain. We consider the challenges of operating system (OS) design targeting next generation high-end computing. We believe that the most urging issues to be addressed are as follows. (1) Exploiting deep memory hierarchies, (2) Reducing cache pollution by OS services and minimizing OS noise, (3) Making it easy to design and deploy application specific kernels, (4) Providing a Linux compatible programming / run-time environment and (5) Enabling seamless tracking of upstream Linux kernel changes. We contend that existing approaches to HPC operating systems, which either employ a stripped down Linux environment or a specific light-weight kernel built from scratch, are not feasible to deal with these challenges. In this paper, we discuss the design decisions of our proposed hybrid kernel design for providing a Linux-compatible light-weight kernel.

著者

Hajime Fujita Yutaka Ishikawa

出版者

Information Processing Society of Japan

雑誌

IPSJ Online Transactions (ISSN:18826660)

巻号頁・発行日

vol.4, pp.73-83, 2011 (Released:2011-03-18)

参考文献数

22

---

In this paper we propose DTS (Distributed TCP Splicing), a new mechanism for performing content-aware TCP connection switching in a broadcast-based single IP address cluster. Broadcast-based design enables each cluster node to continue to provide services to clients even when other nodes in the cluster fail. Each connection request from a client is first distributed among the cluster using the consistent hashing method, in order to share the request inspection workload. Then the connection is transferred to an appropriate node according to the content of the request. DTS is implemented on the Linux kernel module and does not require any modification to the main kernel code, server applications, or client applications. With an 8-node server configuration, a DTS cluster with multiple request inspectors achieves about 3.4 times higher connection throughput compared to the single inspector configuration. A SPECweb 2005 Support benchmark is also conducted with a four node cluster, where DTS reduces the total amount of disk accesses with a locality-aware request distribution and almost halves the number of file downloads that fail to meet the speed requirement.