The deep convolutional network is one of the most successful machine learning models in recent years. However, training large deep networks is a time consuming process. Due to a large number of parameters in these networks, the efficiency of data parallel methods is usually limited by the communication speed of networks. In this paper, we introduce two new algorithms to speedup training large deep networks with multiple machines: (1) propose a new scheduling algorithm to reduce communication delay in gradient transmission and (2) present a new collective algorithm based on reverse-reduce tree to reduce link contentions. We implement our algorithms on a well-known library Caffe and obtain near linearly scaling performance on commodity Ethernet networks.

Null data frames are a special but important type of frames in IEEE 802.11 WLANs. They are widely used in 802.11 WLANs for control purposes such as power management, channel scanning, and association keeping alive. The wide applications of null data frames come from their salient features such as lightweight frame format and implementation flexibility. However, such features can be taken advantage of by malicious attackers to launch a variety of attacks on 802.11 WLANs. In this paper, we identify potential security vulnerabilities in current null data frame applications in 802.11 WLANs. We then study two types of attacks taking advantage of these vulnerabilities in detail that are functionality-based Denial-of-Service attack and implementation-based fingerprinting attack. We also evaluate their effectiveness based on extensive experiments. Furthermore, we design and implement novel defense mechanisms against the attacks, and evaluate their effectiveness based on extensive experiments. Although our proposed defenses help alleviate the vulnerabilities, completely eliminating the vulnerabilities brought by null data frames remains an open issue. Finally, we point out that our work has broader impact in that similar vulnerabilities exist in many other networks due to the adoption of simple and lightweight messages for control purpose.

The teaching environment includes classroom teaching,experiment teaching and network teaching.The construction of teaching environment of elaborate course ‘Structure and Properties of Polymers’ was presented in this paper.

Even though the IEEE 802.11 standards allow multiple non-overlapping channels that can be used simultaneously; most IEEE 802.11-based networks today use only a single channel. Many past researches that attempt to use multiple frequency channels need to modify the 802.11 protocol and therefore can't work well with the 802.11-like networks, and the hidden terminal problem still exists in many proposals. In this paper, we propose one method to utilize multiple channels in wireless mesh networks whose nodes are equipped with multiple network interfaces. The nodes receive instructions on how to use each channel from a centralized algorithm periodically. Following these instructions, the nodes can make full use of each channel and are contention-free. The performance evaluation shows that our algorithm works well in wireless mesh networks.

Null data frames are a special but important type of frames in IEEE 802.11 based wireless local area networks (e.g., 802.11 WLANs). They are widely used for power management, channel scanning and association keeping alive. The wide applications of null data frames come from their salient features such as lightweight frame format and implementation flexibility. However, such features can be taken advantage of by malicious attackers to launch a variety of attacks. In this paper, we identify the potential security vulnerabilities in the current applications of null data frames. We then study two types of attacks taking advantage of these vulnerabilities in detail, and evaluate their effectiveness based on extensive experiments. Finally, we point out that our work has broader impact in that similar vulnerabilities exist in many other networks.

Wireless sensor networks have been widely used in many important areas.Medium access control(MAC) protocols have a significant effect on the function and performance of sensor networks.This paper summarizes the present design paradigms of the MAC protocol in wireless sensor networks.Then,based on the sensor MAC(SMAC) protocol,and by jointly considering the time division multiple access(TDMA),carrier sense multiple access(CSMA) and the synchronization scheme of the SMAC,a new contention-allowed TDMA based MAC protocol for wireless sensor networks(contention-allowed TDMA based SMAC,C-TDMA-SMAC) is proposed.The simulation results show that,compared with SMAC protocol,this protocol has improved a lot in the delay of packets,energy consumption and the reception rate of the packets.

In this paper, we provide a simple, but nevertheless quite precise, analytical model to characterize the short-time unfairness of 802.11 DCF and its effect to the stability of the route in ad hoc networks. We demonstrate that, while pumping packets, the route made up of several wireless nodes acts just like a canal. If the input load is larger than the capacity of the route, conflicts and route broken messages usually generated in the first few nodes. The flow is very stable in the last few nodes. We also show that, for the consideration of route stability, the flow with small rate should be away from the interference range of the congested areas, since, by the result of our calculation, it is hard to get chance capture the channel media.

The interference imposes a potential negative impact on the performance of a wireless network. A device is interfered if it receives a transmission not intended for it. In this paper, we introduce an explicit computation model of node interference based on its actual inducement on the physical layer. We prove that there exists an optimal algorithm to solve the corresponding topology control problem, constructing a network topology with minimum node interference. We also investigate the bad performance of existing methods of solving it and propose a revised algorithm, namely Low Interference-amount Neighborhood Tree (LINT), based on a new link cost metric. The simulation results illustrate that our model is able to reduce node interference effectively and maintain the network performance.