Ryota Kawashima

Non-tunneling Edge-Overlay Model Using OpenFlow for Cloud Datacenter Networks

In current SDN paradigm, an edge-overlay (distributed tunneling) model using L2-in-L3 tunneling protocols, such as VXLAN, has attracted attentions for multi-tenant data center networks. The edge-overlay model can establish rapid-deployment of virtual networks onto existing traditional network facilities, ensure flexible IP/MAC address allocation to VMs, and extend the number of virtual networks regardless of the VLAN ID limitation. However, such model has performance and incompatibility problems on the traditional network environment. For L2 data center networks, this paper proposes a pure software approach that uses Open Flow virtual switches to realize yet another edge-overlay without IP tunneling. Our model leverages a header rewriting method as well as a host-based VLAN ID usage to ensure address space isolation and scalability of the number of virtual networks. In our model, any special hardware equipments like Open Flow hardware switch are not required and only software-based virtual switches and the controller are used. In this paper, we evaluate the performance of the proposed model comparing with the tunneling model using GRE or VXLAN protocol. Our model showed better performance and less CPU usage. In addition, qualitative evaluations of the model are also conducted from a broader perspective.

Scalable and Crash-Tolerant Load Balancing Based on Switch Migration for Multiple Open Flow Controllers

As the size of networks continues to increase, the scalability of the centralized controller becomes increasingly issues in Software Defined Networking. Distributed controllers have been proposed to solve the problem that the centralized controllers such as NOX and Floodlight suffer from. That logically centralized, but physically distributed architecture divide the network into zones with separate multiple controllers to achieve a better scalability control plane. However, Such distributed architecture introduces a new challenge to the load rebalancing of controllers when uneven load distribution among the controllers due to the statically configured mapping between a switch and a controller. Therefore, under variable traffic conditions in real networks, keeping load balance dynamically among the controller clusters is essential for realizing a high performance and scalability control plane. To address these issues, we propose a dynamic load rebalancing method based on switch migration mechanism for clustered controllers. The multiple controllers use Groups to coordinate actions for switch migration. The whole network is divided into several groups and each group is set up one controller cluster. Our proposed method can dynamically shift the load across the multiple controllers through switch migration. The mechanism support controller failover without switch disconnection avoiding the single point of failure problem. We also implemented a prototype system based on Open Day light Hydrogen controller to evaluated the performance of our design. Our preliminary result shows that the method enables controllers to relieve the overload via switch migration and can improve throughput and reduce the response time of the control plane.

SCLP: Segment-oriented Connection-less Protocol for high-performance software tunneling in datacenter networks

The notion of Software-Defined Networking (SDN) has already been introduced into cloud datacenter networks for provisioning virtual network environment. Network virtualization of today is generally achieved by L2-in-L3 tunneling protocols like VXLAN (Virtual eXtensible LAN) and NVGRE (Network Virtualization using Generic Routing Encapsulation) in public cloud datacenters. Some leading production packages for network virtualization have adopted an Edge-Overlay model that performs tunnel encapsulation and decapsulation processes at high-functional virtual switches to utilize existing network equipment. However, a severe performance problem arises because of the software-based tunneling processes. Alternatively, the STT (Stateless Transport Tunneling) protocol overcomes the problem by modifying the semantics of the TCP header, but such changes in semantics raises pragmatic issues in that network middleboxes can discard STT packets as an anomaly. In this paper, we propose a novel layer 4 protocol (Segment-oriented Connection-less Protocol, SCLP) for existing tunneling protocols such as VXLAN and NVGRE. SCLP is designed to not only accelerate the throughput of tunneling protocols, but prevent the packet discarding problem by providing a single-semantic header. Specifically, SCLP can exploit GRO (Generic Receive Offload) feature supported by the Linux kernel to reduce the number of packets to be software-interrupted. We implemented the SCLP protocol and applied it to the VXLAN protocol instead of UDP. As a result, the throughput of the VXLAN over SCLP protocol was almost doubled to the original UDP-based one at maximum.

Implementation and Evaluation of the JobTracker Initiative Task Scheduling on Hadoop

MapReduce is one of the major successful framework to process large-scale data efficiently. Distributed programs can be implemented easily by describing only two methods, Map and Reduce. In Hadoop which is an open source implementation of MapReduce, a JobTracker (master program in Hadoop) assigns Map Tasks and Reduce Tasks to TaskTrackers (slave programs which execute the tasks). In an environment that multiple Hadoops are running on a physical machine, its computational resources should be shared by every Hadoop (Multi-Hadoop environment). In this environment, available computational resources of each Hadoop fluctuate dynamically by behaviors of other Hadoops. Therefore, the JobTracker needs to decide assignment of tasks based on loads and available computation resources on the cluster (JobTracker Initiative Task Scheduler). In this paper, we propose a method which decides the number of task executions in order to use computational resources efficiently based on a load on each computer. And we evaluate its performance, and our results show that the proposal method has achieved a reduction of execution times of jobs by about 11.1% in Multi-Hadoop environment as compared to original Hadoop.

Improving Response Time for Cassandra with Query Scheduling

A management of large-scale data becomes more important, along with the spread of cloud service and the speed-up of networks. Since data management on a single machine can cause performance and scalability problems, data management across multiple machines has been proposed. Distributed Key Value Store(KVS) is a data store which manages data across multiple machines. Since distributed KVSs manage data which consists of simple key-value pair, they can achieve scalability easily. Distributed KVSs are widely used in many services managing large-scale data, such as Facebook and Twitter. Distributed KVSs provide interfaces to access key-value pair by simply specifying the key. In this paper, we refer to a query which only obtains a value from a key as a single query. Some distributed KVSs support a range query which obtains multiple values from a key range. However, under mixed query workloads that consist of single and range queries, single queries(which can be executed faster) are forced to wait until preceding range queries are finished. And this results in the increase of average response time. We propose an approach to reduce the average response time by query scheduling. We implemented our method on Cassandra, and evaluation results showed a reduction of the average response time.

A Host-Based Performance Comparison of 40G NFV Environments Focusing on Packet Processing Architectures and Virtual Switches

Network Functions Virtualization (NFV) is now accepted in large production networks for agile introduction of various Network Functions (NFs). A service chaining technology that dynamically links multiple NFs over the entire network is a heart of the NFV concept. However, virtualizing NF-dedicated hardwares as virtual machines on IA servers brings performance problems, such as lower throughput, longer latency, and larger jitter. Various packet processing frameworks and virtual switches have been proposed to resolve the problems, but there is no comprehensive study of their performance characteristics focusing on both physical/virtual layers. In this paper, we evaluate fundamental throughput and latency/jitter of three packet processing architectures (NAPI, netmap, and DPDK) with six virtual switches (Linux Bridge, Open vSwitch, VALE, L2FWD-DPDK, OVS-DPDK, and Lagopus) for physical/virtual layers. Our experiments were performed on both Intel and Mellanox 40 GbE NICs. Finally, we discuss appropriate NFV host environment for commercial-use.

Implementation and Performance Analysis of STT Tunneling Using vNIC Offloading Framework (CVSW)

Network Virtualization Overlays (NVO3) provides multi-tenancy services in cloud data centers with existing networking equipment. IP tunneling is an essential technology to logically separate each virtual traffic, in particular, Stateless Transport Tunneling (STT) is considered to achieve better performance using TCP Segmentation Offload (TSO) feature. Currently, there is no openly available implementation of STT, and its implementation and performance characteristics have not been studied in academic field so far. We have therefore implemented STT protocol and conducted performance evaluation by comparing with VXLAN protocol. In practice, the STT implementation has been done using a virtual NIC offloading framework, co-virtual switch (CVSW). CVSW is a software component that extends virtual NICs and provides high-level packet processing framework such as Open Flow Match-Action. In this paper, we describe implementation details of STT and performance evaluation results from various perspectives. The results showed that the actual performance of STT was almost equal to non-tunneling VM-to-VM communication and was two-times higher than that of VXLAN. Furthermore, we clarify the high-performance nature of STT is brought from both byte-stream characteristic of TCP and Generic Receive Offload (GRO) feature rather than widely believed TSO.

vNFC: A Virtual Networking Function Container for SDN-Enabled Virtual Networks

Software-defined networks (SDN) has gradually been deployed on commercial networks such as datacenter networks. Current SDN is based on OpenFlow technology that is a set ofnetwork flow control API for switch devices. For instance, network reachability between end-hosts (or virtual machines), packet filtering mechanisms, and status management of switches are enabled by the API. In practice, however, current OpenFlow-based SDN has following problems: no application-layer protocol support and switch-oriented flow control. Since OpenFlow targets L2-L4 flow handling, users have to arrange additional mechanism for upper-layer flow control. Furthermore, executing a lot of flow matching on a single switch (or virtual switch) can cause difficulty in network trace and overall performance degradation.This paper proposes a virtual networking function container (vNFC) that is a set of software implemented networking functions for VM-to-VM communications, and it is located between a virtual machine and a virtual network device of the host machine. vNFC enables not only lower-layer functions OpenFlow providing, but also upper-layer functions like application firewall in the same manner. That is, vNFC is a virtual machine dedicated flow handling function set. In addition, OpenFlow-compatible vNFC configuration protocol named OpenNF and vNFC controller are also presented. OpenNF provides communication path between each networking function and the controller for configuration and decision making.In this paper, architectural design and implementation of vNFC are presented, and also performance evaluation of using vNFC. The evaluation result shows that a lightweight networking function does not impact on the performance, but a function that frequently communicates with the controller incurs millisecond order cost per frame transmission.

Evaluation of Forwarding Efficiency in NFV-Nodes Toward Predictable Service Chain Performance

The concept of network functions virtualization (NFV) has been embodied in commercial networks over the past years. Software-based virtual network functions have forwarding performance concerns in general, and various acceleration technologies have been developed so far, such as DPDK and vhost-user. Existence of several alternatives requires network engineers or operators to select appropriate technologies; however, no pragmatic criterion exists for constructing high-performance NFV-nodes. From their points of view, a lack of common benchmark and understanding of performance characteristics makes it difficult to predict hop-by-hop performance in a service chain, which results in prevention of NFV deployment in mission-critical networks. In this paper, we clarify performance characteristics of packet forwarding in NFV nodes focusing on three types of acceleration technologies; packet I/O architecture, virtual network I/O, and forwarding engine in a practical stage. We examined three packet I/O architectures (NAPI, netmap, and DPDK), three virtual I/O mechanisms (vhost-net, vhost-user, and SR-IOV), and four practical forwarding programs (Open vSwitch, OVS-DPDK, xDPd-DPDK, and Lagopus) with three referential programs (Linux Bridge, VALE, and L2FWD-DPDK). The experiment was conducted on a 40 GbE environment and we examined two device-under-test machines having different CPU performance. We argue performance characteristics of each technology and give quantitative analyses of the result. The key findings are: 1) CPU core speed has impact on both throughput and latency/jitter; 2) DPDK can allow performance prediction; 3) vhost-user is appropriate for real environment; and 4) OVS-DPDK provides a good combination of performance and functionality.

Deferred-Update Replication Supporting Multiple Consistency Models

Many distributed systems use a replication mechanism for reliability and availability. On the other hand, developers have to consider minimum consistency requirement for each application. Therefore, a novel replication protocol that supports multiple consistency models is required. Multi-Consistency Data Replication (McRep) is a middleware-based replication protocol and can support multiple consistency models. However, McRep has a potential problem in that a replicator relaying all request and reply messages between clients and replicas can be a performance bottleneck and a single-point-of-failure. A deferred-update replication is a well-established approach for fault-tolerant data management systems but can only guarantee One-Copy Serializability. We apply the version control method of McRep to the deferred-update replication protocol to solve the problems. We extend client-side programs to store version numbers of their own transactions. We also extend the replicas role to control version numbers to determine if a replica ensures a specified consistency model. We have implemented and evaluated both the proposed and the McRep protocols. The evaluation results show that the proposed protocol achieved comparable throughput of transactions to McRep. Especially the proposed protocol improved the throughput up to 16% at a read-heavy workload in One-Copy Serializability.