Hardware support for reliable networking; from protocols to accelerators

点击翻译按钮获取中文摘要

The long-term goal of this research program, toward which significant contributions are expected over the next 5 years is to obtain reliable and effective electronic systems. Several threats and failure mechanisms can jeopardize reliability of electronic systems and these systems are effective when they provide the right functionality, while offering adequate throughput, latency, power dissipation and energy consumption. If reliability and effectiveness/efficiency can be orthogonal concerns, the proposed research aims at contributing to both to get balanced and better systems. A growing concern is the reliability and efficiency of datacenters that notably support cloud services. Key issues stem from to the steady growth of link data rates (commonly 40 Gb/s, 100 Gb/s emerging, 400 Gb/s and 1 Tb/s on the horizon). With growing rates, the time to process packets shrinks and two reliability issues must be tackled: transmission errors and distributed denial-of-service attacks (DDoS). Ineed, a very significant and pervasive threat to computer systems from the Internet is the DDoS. The goal of a DDoS is to saturate a computer resource thus making it unavailable to legitimate users. The proposed program will first propose protocols and flow control to improve datacenter network reliability. One protocol will support link level integrity checking and retransmission. The program will then investigate mappings between packet function and bandwidth usage to support passive defense against DDoS. Also, we will devise and implement algorithms to support this mapping concept to provide resource throttling as a resilient defense to protect against DDoS. These algorithms will be distributed in nature; otherwise, the filtering of packets would be an attackable service. In other words, the effort to defend could be so large that concentrating it at the attacked service could still overwhelm the rest of the system. The result will be that resources (i.e. services) will be protected against DDoS attacks intrinsically by the behavior of the network. On the other hand, this defense does not differentiate legitimate traffic from an attack thus the resulting throttling would impact both traffics. We propose to further improve the defense mechanisms with detection and mitigation of attacks based on techniques like principal components analysis and deep learning. To the best of our knowledge, providing a passive defense infrastructure against has never been proposed in the literature. The program will also aim at improving datacenter network reliability through protocols and flow control, and to extend these protocols to enable cyber attack mitigation, and to propose and validate means of effectively detecting cyber attacks.