Yukinobu Fukushima

A reroute method to recover fast from network failure

The demand for high Internet availability has increased. As routing methods to recover fast from a single network component (link or node) failure, methods using multiple routing tables have been proposed. In the methods, when a component failure occurs, packets which are supposed to pass through the component in the failure-free state are rerouted using a routing table (backup table) which does not use the component for packet forwarding. Thus, it is important to decrease the increased routing cost (for example, the number of increased hops of rerouted packets if every link cost is one) due to the reroute as much as possible. In this paper, we propose a new routing method to decrease the increased routing cost. In the conventional methods, there are many components that do not fail but are not used for packet forwarding in a backup table. In our proposed method, we use such components as much as possible, and consequently we try to decrease the increased routing cost. Numerical examples show the effectiveness of our proposed method.

Performance evaluation of as-friendly peer selection algorithms for P2P live streaming

In this paper, we evaluate the performance of peer selection algorithms: MLH (Minimum Logical Hop) and MPH (Minimum Physical Hop) to increase the number of joining peers and to decrease inter-AS traffic volume in P2P live streaming, where we assume that every peers logical hop count (the number of providing peers between an origin streaming server and the peer) is limited by a predetermined value in order to keep real-time property of live streaming. In MLH, a newly joining peer selects such providing peers that the logical hop count is minimum to increase the number of joining peers. And then if there are several such providing peers, the newly joining peer selects such providing peers that the number (physical hop count) of ASs between the newly joining peer and each of the providing peers is minimum to decrease the inter-AS traffic volume. In MPH, a newly joining peer selects providing peers in the reverse order of MLH. Simulation results show that MLH achieves about from 20% to 60% more maximum joining peers than MPH, and MPH shows smaller inter-AS traffic volume than MLH when the number of joining peers is small while MLH shows smaller inter-AS traffic volume when the number of joining peers is large.

Performance evaluation of a multi-stage network event detection scheme against DDoS attacks

Change-point detection schemes, which represent one type of anomaly detection schemes, are a promising approach for detecting network anomalies, such as attacks and epidemics by unknown viruses and worms. These events are detected as change-points. However, they generally also detect false-positive change-points caused by other events, such as hardware problems. Therefore there is a requirement for a scheme that detects only true-positive change-points caused by attacks and epidemics by unknown viruses and worms. The true-positive change-points tend to occur simultaneously and intensively in very large numbers, while the false-positive change-points tend to occur independently. We can exclude false-positive change-points by excluding those that occur independently, based on information gathered from the entire network. In this paper, we combine change-point detection schemes with a distributed IDS, and evaluate performance of the combined scheme by a simulation using the parameter values obtained by an experiment using real worms. The simulation results show that the combined scheme detects all the DDoS attacks without any false-positives while we have to tolerate false-positive rate of at least 0.02 to detect all the attacks in a stand-alone IDS scheme.

Improvement of a TCP Incast avoidance method for data center networks

In distributed file systems, a well-known congestion collapse called TCP Incast occurs because many servers send data to the same client and then many packets overflow the port buffer of the client link. Incast leads to throughput degradation in the network. In our previous work, we have proposed a method to avoid Incast. In the method, we limit the maximum number of simultaneously existing connections to a predetermined constant value. However, we cannot use the method when the data size is small, and because we have not investigated how to optimize the maximum value, Incast may occur if the maximum value is not appropriate. In this paper, we improve the method so that it is applicable regardless of the data size and the maximum value is optimized. Numerical results show the effectiveness of our proposed method.

Optimization of Token Holding Times in Split Light Trail Networks

As a new optical WDM network architecture that can be built with currently available devices and can achieve bandwidth allocation with granularity finer than a wavelength, a light trail architecture attracts attention. Because a light trail is a shared medium, we need a medium access control (MAC) protocol to avoid collisions. Although MAC protocols using token passing can avoid collisions, bandwidths of links that locate upstream of the token holding node are kept idle. In this paper, we first propose a dynamic light trail splitting method in order to increase throughput of a light trail by using those idle bandwidths. Our method splits a trail into the upstream trail and the downstream trail at the token holding node and independent data transmissions on the two trails are permitted. As a result, we expect that the split trail architecture achieves higher maximum throughput than the original non-split trail architecture. The degree of throughput improvement by the split trail architecture depends on how appropriately we set upstream and downstream token holding times of every transmission node. Thus, we formulate a problem to decide the token holding times as a nonlinear programming problem, derive the maximum throughput of the split trail architecture by solving the problem using NUOPT solver, and investigate the degree of improvement compared to the original architecture. According to numerical examples, the split trail architecture achieves 1) almost the same maximum throughput as the original one for its unfavorite traffic pattern where every transmission node sends data to the terminating node of the trail only, 2) about 1.6 times as high maximum throughput for a uniform traffic pattern where every node-pair requests the same traffic volume, and 3) about 1.9 time as high maximum throughput for its favorite traffic pattern where every transmission node sends data to its adjacent downstream node only.

Throughput Optimization in TCP with a Performance Enhancing Proxy

To improve TCP throughput performance, a method using a PEP (Performance Enhancing Proxy) has been proposed. The PEP operates on a router along a TCP connection. When a data packet arrives at the PEP, it forwards the packet to the destination host, transmits the corresponding ACK (premature ACK) to the source host in behalf of the destination host, and stores a copy of the packet into its own buffer (PEP buffer) in case retransmission of the packet is required. As a congestion control method on the PEP, a method which keeps the number of prematurely acknowledged packets in the PEP buffer below a threshold (watermark) value has been proposed. However, the relation between the watermark value and throughput is not sufficiently investigated, and an optimization method of the watermark value is not proposed. In this paper, we first investigate the relation between the watermark value and the average throughput. Extensive simulations show that the simulation results are roughly classified into two cases. In the first case, the average throughput becomes larger for larger watermark values and becomes a constant value when the watermark is over a certain value. In the second case, although the average throughput becomes larger for larger watermark values in the same way, it decreases when the watermark is over a certain value. Next, based on the results about the relation, we propose an watermark optimization algorithm which can adaptively maximize the average throughput of each connection and also satisfy a fairness condition that the average throughputs of connections are equal to each other.

TCP incast avoidance based on connection serialization in data center networks

In distributed file systems, a well-known congestion collapse called TCP Incast occurs because many servers send data to the same client and then many packets overflow the port buffer of the client link. Incast leads to throughput degradation in the network. In this paper, we propose two methods to avoid Incast based on the fact that only one connection can almost fully use the network bandwidth because the bandwidth-delay product is not so large in current data center networks. The first method is a method which completely serializes connection establishments. By the serialization, the number of packets in the port buffer becomes very small, which leads to Incast avoidance. The second method is a method which overlaps the slow start period of the next connection to the current established connection to improve throughput in the first method. Numerical results from extensive simulation runs show the effectiveness of our two proposed methods.

An on-line algorithm to determine the location of the server in a server migration service

In IaaS cloud services, QoS of network applications (NW-Apps) may degrade due to location factors such as significant distance between a server-side application (server) of a NW-App at a data center and a client-side application (client) of the NW-App at a client terminal. In order to shorten the distance and to improve the QoS, server migration services (SMSes) have been proposed. In SMSes, servers may migrate between different computers (called work places, WPs) on a network to prevent QoS degradation caused by the changes of client locations. Although server migrations can improve QoS of NW-Apps, they also generate a huge amount of traffic (server migration traffic) in the network. This paper focuses on a server location decision problem where the location of a server is decided in an on-line manner so that QoS of a NW-App is improved under the constraint that the server migration traffic has to be suppressed below an acceptable level. For the problem, we propose a practical on-line algorithm. The key idea behind the proposed algorithm is that the location of the server is decided with consideration of the QoS degradation in the future. The algorithm defines the averagely good location for the server where the QoS is expected to be relatively good for various client locations. Then, it keeps the range of the servers migration within the returnable range where the server can soon come back to the averagely good location. As a result, the QoS can be always kept as good as the one under the averagely good location. Simulation results show that the proposed algorithm improves QoS of the NW-App by up to 30% compared to a greedy algorithm.

A trail multi-splitting method for throughput improvement in light trail networks

Light trail networks have been proposed as an extension of lightpath networks in order to enable sub-wavelength bandwidth allocation in WDM networks. The networks need a medium access control (MAC) method to avoid collisions of data transmissions. We have proposed a MAC method which adopts a token-passing mechanism. The method splits one trail into two trails: one upstream trail and one downstream trail and permits independent data transmissions on the trails. However, the method leaves bandwidths of links upstream from the token-holding node idle because the method only splits the original trail into two trails and leaves the upstream trail unsplit. In this paper, we propose a novel token-passing MAC method for further improving trail throughputs. Our proposed method recursively splits the upstream trail, and consequently the original trail can be split into more than two trails. In the proposed method, an appropriate setting of upstream/downstream token-holding times of all nodes is a key factor to successfully accommodate an input traffic. Therefore, we tackle a problem (called a token-holding time decision problem) where we optimize token-holding times of all nodes so that input traffic is successfully accommodated. We formulate the problem as a linear programming model. Numerical examples on maximum effective throughput show that the proposed method is more superior as the ratio of short-hop traffic is higher in input traffic patterns.

Optimization of maximum timeout value in TCP with a fine-grained timer for data center networks

When the standard TCP implementation is used in distributed file systems in data center, a well-known throughput degradation called TCP Incast occurs. In order to avoid TCP Incast, a TCP implementation using a specialized fine-grained timer is proposed. However, the throughput still decreases in the implementation when the number of servers in distributed file systems is large. In this paper, we clarify that the cause of the decreasing is that maximum timeout value is too large and optimize the value.