Computer & Software Engineering Group Seminars Archive

2007 May 4, Scott Miller: Trends in Low-Power FPGA Interconnection Network Design

Abstract:
With increased interest in reconfigurable computing for rapid-prototyping and augmenting microprocessor-based systems, FPGAs are becoming increasingly attractive solutions to these applications. It is well known that the interconnection network of an FPGA consumes up to 90% of silicon-resources, and accounts for up to 70% of the energy consumed on-chip.

In this talk, trends in design methodologies for low-power FPGA interconnection networks will be addressed. Topics such as programmable switch architectures and sizings, buffer design, and line driver design and sizings will be addressed. These topics will also be discussed according to their trends across multiple CMOS IC fabrication technologies. Energy consumption, delay, area and noise immunity concerns will be presented and design methodologies will be compared based on these metrics. Simulation results will present data on optimal component sizes with respect to a composite area-delay metric.

2007 March 23, Dr. Wei LU: NetShield - An Adaptive DDoS Detection and Mitigation System

Abstract:
With more than a million hosts already controlled by hackers, Distributed Denial of Services (DDoS) attacks by Botnets are becoming a severe threat on current web information systems. Traditional approaches against DDoS attacks on the victim end tend to generate high false alarm rate and thus lead to more collateral damage on quality of web services.
In this talk, NetShield, an adaptive DDoS detection and mitigation system, is presented in order to address limitations of current DDoS defense approaches. NetShield system is composed of two components: detection module and mitigation module. In detection module, CUSUM algorithm and EM based clustering algorithm are applied to identify DDoS intrusions in realtime; In mitigation module, filtering rules are set based on historical web access and are activated once the front detection module triggers DDoS alarms. On a lab evaluation with German Telecom, a Botnet with 100 attackers was simulated to generate the highly structured DDoS traffic and the stand-alone NetShield system successfully detected and blocked all malicious traffic to the target server in a short time.

2006 December 8, Kaveh Aasaraai: Low-Power Perceptron Branch Predictor

Abstract:

Branch predictors relying on neural networks have received increasing attention in recent years. Unfortunately, such designs are often impractical as they come with high latency and power dissipation. In this work we introduce a low-power perceptron predictor which utilizes as much resources as needed according to the branch behavior, effectively reducing overall number of computations. We reduce predictor energy consumption by not assigning computation resources to unnecessary computations. While reducing predictor energy consumption, we also improve overall performance as we reduce prediction latency. We reduce the predictor computational power dissipation up to 34% while improving the processor performance by up to 19%.

2006 December 1, Muhammad Nadzir Marsono: Spam Control Based on Layer-3 Classification

Abstract:
Spam constitutes approximately two-thirds of the e-mail traffic over the Internet and is fast becoming a major problem for both users and administrators. Spam reduces performance of network infrastructures, frustrates e-mail users, and costs billions in lost productivity. In recent years, spam was also relayed by worm-infected hosts that work as spam delivery agents. Existing spam control systems work on fully reassembled (layer-7) e-mails, after they have been received, on e-mail servers. This is costly since e-mail servers are not protected against attack. If the processing is distributed over the network nodes, the cost is distributed amongst many Internet users. Processing e-mails at layer-7 is too complex and too slow to be performed on intermediate nodes.

In the first part of this talk, we propose a distributed layer-3 e-mail classification approach. E-mails are classified at the packet level, which does not require reassembly. Thus requiring less computation overhead and suitable for high throughput hardware implementations. This makes distributed e-mail classification on upstream network nodes possible. We focus on the functional accuracy of the layer-3 e-mail classification as opposed to the layer-7 using modified naive Bayes classification technique. We show how the naive Bayes technique can be adapted to support pre-classification at layer-3 and simple aggregation at layer-7. Our results, based on simulations with TREC2005 corpus, show that legitimate e-mail misclassification is approximately equal to the layer-7 accuracy, and even better in small packet lengths. However, classifying e-mails at the packet level is still susceptible to evasion and insertion attacks. Instead of using the layer-3 e-mail classification information as a final decision either to accept or to reject an e-mail, it is used to predict e-mail classes for better e-mail services on e-mail servers.

In the second part of this talk, we analyze the performance of three proposed e-mail service strategies on e-mail servers by utilizing the layer-3 e-mail classification. Instead of using a single-queue system, where non-spam and spam alike are process fairly, a two-queue system is proposed instead. We model these strategies using discrete-time Markov chain analysis. The behavior of the priority queue is evaluated under different e-mail traffic (non-spam and spam) loads and service capacities. Our models show that non-spam e-mails can be delivered under small delay even under heavy e-mail loading and high spam to non-spam prior. The loss probability of non-spam e-mail mistakenly forwarded to the slow queue is smaller than the one for a single queue service strategy. Servicing e-mails by predicting whether or not an e-mail is a spam has two-fold advantages. First speeding up non-spam e-mails services and delivery while delaying spam e-mails, and second protecting e-mail servers in the events of attacks.

Processing e-mail at layer-7 requires complex processing, which restricting deployment on e-mail server. We proposed the layer-3 classification to reduce the processing complexity.The layer-3 e-mail processing enables distributed processing on network nodes. By doing some form of processing to e-mail packets in-transit, better service strategies can be performed on e-mail servers. This increases the service performance on e-mail servers to support the future e-mail traffic increase.

2006 November 24, Farshad Khunjush: Chip Multiprocessing and the Cell Broadband Engine