Nobutaka Matsumoto

LightFlow: Speeding up GPU-based flow switching and facilitating maintenance of flow table

Flow-based switching is increasingly important in accordance with the growing demand for in-network processing for cloud applications. Flow switching performance tends to be degraded in proportion to the number of flow entries. To reduce the number of flow entries, they can be aggregated by applying wildcard fields. Meanwhile, the existence of the wildcard entry adversely affects the use of a hash-based lookup on a flow table, and thus a linear search is inherent in flow switching. However, the linear search is currently the primary cause of performance limitation. To date, two flow tables, one for hash-based lookup and the other for a wildcard-enabled linear search, have been used for flow switching. While hash-based table lookup is much faster than linear search, it needs to be manually updated for every exact match entry. Maintaining a hash-based table of all the flow switches is not feasible from a network operator viewpoint. In this paper, LightFlow, a mechanism to accelerate software flow switching processing and relieve the burden of maintaining the flow table is proposed. In LightFlow, two-dimensional parallelization of a linear search is introduced to accelerate lookup of the wildcard-enabled flow entries. It also introduces a mechanism that allows updating of the hash table to be performed automatically based on the result of wildcard-aware table lookup. LightFlow satisfies both the need for fast table lookup and feasibility of flow table management which needs to allow a large number of wildcard entries. Experimental results show that LightFlow can increase the speed of lookup of a wildcard-aware flow table three-fold or more compared to the current GPU-based wildcard search mechanisms.

Orchestration of heterogeneous virtualized resources for end-to-end service control

To achieve end to end (E2E) quality management capability for different services (e.g. voice, video and data), an SOA-based E2E resource orchestration architecture based on hierarchical distributed model and virtualization mechanisms for individual networks is proposed. Possible implementations and operation mechanisms considering possible service scenarios are discussed. Prototype of resource management (RM) system is demonstrated for the first time to manage and control distributed GE-PON-based access network resources, RPR-based metro network resources and GMPLS-based core network resources. Proposed RM system consists of network resource managers (NRMs) virtualizing individual network resources and network resource broker (NRB) leading orchestration of NRMs with a single point of contact to a service control layer. Exposure of E2E network service to a service control layer as well as the exposure of virtualized network services to NRB are represented by a universal Web services interface (WSI). Using the RM system, the heterogeneous resources are orchestrated to provide appropriate bandwidth-managed E2E services following service demand of a network service client, and appropriate resources are applied to assumed voice, video and data flows.

Network Middleware Design for Bridging Legacy Infrastructures and NGN

Session-based QoS control mechanism is a key characteristic in next generation networks (NGN). However, the session-based QoS control defined in NGN standardizations is not designed for layer 2 infrastructures, but for border gateway function (BGF) at IP edge. To achieve QoS assurance for NGN services, end-to-end network resource management, including layer 2 access and metro networks, is required. Although the BGF function needs to be introduced at IP edge, legacy infrastructures must be utilized especially for access and metro networks. Incorporating those widely deployed legacy infrastructures into session-based QoS control mechanisms is expected. However, differences in user-flow identification of NGN and resource identification required for managing legacy infrastructures induce gaps blocking integration to the unified session-based infrastructure. This paper proposes to use a network middleware for bridging NGN control frameworks and legacy infrastructures. With the multiple service oriented architecture (SOA) chain, the network middleware distributedly coordinates middleware element managing each network segments, and harmonizes differences of resource description and control procedures in NGN and legacy infrastructures. The network middleware enables to maintain coherent QoS policy for the end-to-end route of the session-based services throughout the heterogeneous and multi-layer infrastructures.

GPU-accelerated hash and wildcard hybrid flow switching for tackling massive flow entries

GPU-accelerated high performance switching design using hybrid flow tables allocated in host and GPU memories is proposed based on LightFlow architecture. In this work, careful placement of hash and wildcard tables is proposed considering overhead of data copy during lookup process, in order to tackle massive flow entries using commodity hardware. Experimental results show that the proposal achieves 14.6 Mpps which is the fastest record of the software flow switch operated under a large scale reaching 4 million flow entries. With this proposal, it is also demonstrated that the average processing delay is about one forth compared with Open vSwitch.

BPEL Driven Policy Management of Virtualized Network Resources for IMS Environments

BPEL Driven Policy Manager (BDPM) designed to coordinate network resources with Web service interfaces for IMS environments is developed. BPEL workflow is applied to the policy manager for achieving end-to-end assurances of QoS and bandwidth on heterogeneous network resources. BDPM analyzes QoS-related information from IMS, and identifies appropriate policy workflow to invoke end-to-end network services by allocating resources adaptively. The resource allocation results are enforced to the distributed network-resource-management systems using service oriented architecture (SOA) through Web service interface. Using prototype implementation of BDPM, various descriptions of QoS requirements from IMS core can be adaptively processed by BDPM and end-to-end network resource management is successfully conducted.

Erlang-k-based packet latency prediction model for optimal configuration of software routers

Providing the optimal configuration for a software router poses a lot of technical challenges that do not present in the dedicated hardware router. One of them is how to characterize performance varying due to different configurations on commodity hardware. This paper addresses the problem of configuring a software router that provides the minimum of average packet latency. Since changing all combinations of hardware configurations of a software router for searching the optimum is cumbersome, we propose a prediction model to accurately estimate the packet latency of a software router. We first analyze the relationship of the packet latency distribution with the configured and observed parameters. Empirical measurements suggest that the Erlang-k distribution is a reasonable model for estimating the packet latency distribution. Motivated by the parameter relationship analysis, we propose a prediction model for packet latency of a software router based on the Erlang- k distribution. Our prediction model requires measurement of only two different configurations, i.e., one and two Rx queues of a network interface card, to predict the average packet latency of all combinations of configurations. We use the measured data from the testbed experiments and the data of curve fitting method to cross-verify the accuracy of our prediction model. Underlying the prediction model, we propose the optimal configuration selection (OCS) algorithm to justify which configuration yields the minimum of average packet latency. Our prediction model based OCS results in the same optimal configuration with the measured data based ones.

Roles of network federations for bridging multi-generation standard architectures

Towards the future development of global ICT environments, roles of network federation will become increasingly important and complex. Functional gaps between standard network architectures, such as conventional IP-based networks, next generation networks, and future Internet, tends to be remarkable. Architectural migrations from conventional IP to next generation network and to future Internet are analyzed to identify the role and remaining issues of the network federation of different standardized architectures. Possible solutions to network federation are proposed to address migrations to the next generation network and future Internet.

A demonstration of session-based resource reservation interface for inter-NGN QoS control

The inter-domain QoS control employing a session-based resource management capability of next generation networks (NGN) is demonstrated for the first time. For a video-on-demand application, the extended resource and admission control function (RACF) is applied and the resource reservation over different NGN-domains is successfully conducted.