川流處理架構之高速網路封包處理

近年來，網路快速的發展，已使人們對相關的網路應用與需求，從傳統路由、及管理，轉移到需要更深入封包內層檢查，或是更改其內容衍生之議題，諸如基於內容計費、服務品質、第七層的路由交換、以及網路安全等相關的應用。然而這些應用都有些共同的特色：需要處理川流不息的封包中每一個位元組的資料，而且這些額外的處理工作，已經超出原有網路系統設備所能負荷之處理能量。傳統的網路系統實現技術使用一般微處理器架構，但因網路系統運算的複雜度，使用一般低效率封包處理能量的傳統微處理器架構，很難達到硬體線速(wire-speed)的效能。因此目前大部份的網路系統設計，是使用特定用途晶片(ASIC)來達到高速流量處理的目標，但其固定架構的特色無法滿足網際網路通訊協定快速變化，和系統設計彈性化的需求。於是、工程師們使用網路處理器(Network Processors)，以其優化之多核、多執行緒處理器架構、達到一個結合特定用途晶片(ASIC)在效能上的優勢，與一般微處理器可程式化彈性的優點，來達成高速網路系統設計的目標。

川流處理架構原為提供影像處理所需之高效能數位訊號(DSP)處理器。在封包處理的過程中與影像處理有很多如：(1)高運算複雜度與、(2)高資料平行度，等相同之特色。此外，與多執行緒架構比較起來，川流處理架構沒有像多執行序微處理器中，要具備有記錄每一個執行緒系統狀態的控制、與記憶電路，也因此節省了大量的系統晶片面積與功率消耗。所以，川流處理架構已被應用為另一具備高速網路封包處理，與程式化及彈性兼具的架構。

My research is focused on computer networks with particular interests on integrating novel architecture & functionality into network systems. We work on topics related to embedded design for network security and traffic measurement on FPGA and Network Processors.

High-Speed Packet Processing & Measurement

Traffic analysis and measurement are important tasks for the proper operation of IP networks. The accurate estimation of Internet traffic statistics serves as the basis for infrastructure planning, network provisioning, capacity forecasting and accounting. Anomaly detection on worm distribution and prevention of distributed denial of service (DDoS) attacks are also based on the same information. However, as network bandwidth grows exponentially, the scaling of monitoring and measuring capabilities for collecting accurate statistics becomes a critical issue. Hash-based algorithms are very useful and popular techniques adopted in many high-speed router design. We are exploring these advanced techniques with hardware and architecture support for data reduction and synopsis construction.

Packet Processing on Stream Architecture

Stream processing architectures have been proposed as efficient and flexible platforms for network packet processing. This is because packet processing shares many of the same characteristics of media and image processing that motivate stream architectures: little global data reuse, abundant data parallelism, and high computational complexity. Moreover, in comparison with multithread approach, stream architecture provides much lower cost to hide a given amount of latency.