Osamu Shiraki

AP1000+: architectural support of PUT/GET interface for parallelizing compiler

The scalability of distributed-memory parallel computers makes them attractive candidates for solving large-scale problems. New languages, such as HPF, FortranD, and VPP Fortran, have been developed to enable existing software to be easily ported to such machines. Many distributed-memory parallel computers have been built, but none of them support the mechanisms required by such languages. We studied the mechanisms required by parallelizing compilers and proposed a new architecture to support them. Based on this proposed architecture, we developed a new distributed-memory parallel computer, the AP1000+, which is an enhanced version of the AP1000. Using scientific applications in VPP Fortran and C, such as NAS parallel benchmarks, we simulated the performance of the AP1000+.

DomainFlow: practical flow management method using multiple flow tables in commodity switches

A scalable network with high bisection bandwidth and high availability requires efficient use of the multiple paths between pairs of end hosts. OpenFlow is an innovative technology and enables fine-grained, flow level control of Ethernet switching. However, the flow table structure defined by OpenFlow is not hardware friendly and the scalability is limited by the switch device. OpenFlow is also not sufficient for fast multipath failover. To overcome these limitations, we propose DomainFlow in which the network is split into sections and exact matches are used where possible to enable practical flow management using OpenFlow for commodity switches. We applied a prototype of DomainFlow to multipath flow management in the Virtual eXtensible LAN (VXLAN) overlay network environment. The total number of flow entries was reduced to 1/128 using currently available commodity switches, which was not possible before.

A Single-Chip, 10-Gigabit Ethernet Switch LSI for Energy-Efficient Blade Servers

The use of virtualization technology has been increasing in the IT industry to consolidate servers and reduce power consumption significantly. As a virtualization platform, a large-scale blade server is suitable because it can hold a dozen blades in a chassis with well managed configuration, enabling easy provisioning. To realize an energy-efficient blade server, the network component must deliver both high performance and reduced power consumption. We developed the fifth generation single-chip 10GbE switch LSI that supports 26 10GbE ports with built-in 10 Gb/s serial back plane interfaces. Using this highly integrated switch LSI, we also developed a single-wide 10GbE switch blade for the blade server. The switch blade delivers 100 percent more performance per watt than other 10GbE switch blades in the industry. This paper describes the features of the switch LSI, the high-speed IO circuit of its built-in interfaces and 10GbE switch blade.

Dynamic virtual network configuration between containers using physical switch functions for NFV infrastructure

Container-based virtualization simplifies the deployment of applications and will have a significant impact on the future of data center networks. In an NFV infrastructure, it is required to dynamically extend/reduce resources granted to the virtual network function (VNF) as needed. In this paper, we explored SDN technology to construct lightweight virtual networks between Linux containers in dynamic container placement to isolate tenant traffic. We used an SDN controller to distribute logical endpoint information to physical switches before the host physical location was determined for the container. And we used physical switch functions to promptly configure virtual networks between containers when a container was dynamically placed on a host and started in a split second. We also actively probed the container and automatically removed the virtual network when the container stopped. By using our prototype, virtual networks were automatically constructed with no interaction with the SDN controller as soon as 500 of the containers were started.

Managing storage flows with SDN approach in I/O converged networks

The demand for efficient use of information and communication technology (ICT) resources has recently been growing. Equipment costs at datacenters have been reduced to satisfy this demand and all ICT resources have been required to be controlled on demand. Network virtualization and software defined networking (SDN) technologies have been applied to datacenter networks for the same reason. On the other hand, efficient networks and operations also need to be provided, including traditional storage systems. Ethernet fabric and I/O convergence have been applied to satisfy these requirements. However, the SDN approach has mainly been applied to local area networks and flexible control has so far not been accomplished for converged networks. We introduced an SDN approach similar to OpenFlow and applied this to Fibre Channel over Ethernet (FCoE) networks, which is a method of achieving I/O converged networks. This method of controlling flows can coexist with OpenFlow based SDN flow control. We explain the differences between our method of flow control, OpenFlow, and path optimization with FCoE and their effects as examples in this paper.

Flow-Aware Congestion Control to Improve Throughput under TCP Incast in Datacenter Networks

Data enter networks (DCNs) host cloud computing and various applications. In this environment, there are mice and elephant flows and the partition/aggregate communication pattern is common, so TCP in cast may often occur. Massive mice flows directly affect the elephant flows under in cast, and the standard TCP mechanism or SDN approach with centralized controller is insufficient to solve this problem. In this paper, we focused on how a flow-aware congestion control method that takes into account flow characteristics can improve throughput under in cast. It is required to establish our long-term goal that provides different network level services based on flow characteristics. We propose a switch-based method that identifies target flows and provides different congestion control using explicit congestion notification (ECN). Although our method uses ECN, it does not require any change to TCP stacks on end nodes or a dedicated queue for the targets. In our experiments, we focused on the throughput of flows, and selected elephant flows as targets because their throughput is more sensitive than mice flows. We compared the experimental results from our method to those of standard TCP, ECN, and a dedicated queue for elephants. Our method improved the aggregate throughput of elephant flows by 1.7× to 5.1× and the good put of elephant flows by 1.9× to 5.4× compared to those of the above methods under our in cast scenario.