Yao Yue

System Core Group, System Software Department, NVIDIA, Santa Clara, CA
Internship in CUDA-Enhanced Compression for Remote Graphics

Motive: As GPU becoming the new computing powerhouse, GPU virtualization can allocate dedicated GPU resources to multiple clients for concurrent, high quality display. For remote clients, dozens of frames will be delivered over the network in a second, easily generating astronomical traffic volume (several Gbps). Compression is thus a must for remote graphics. However, CPU-based compression mechanisms in today¡¯s remote desktop fall short of the requirements of many applications. Further, algorithms chosen should adapt to varying tradeoffs between image sharpness and display smoothness.

Goal: Design and implement a GPU-based software compression module that converts screenshots into a compact stream ready for network transmission, and meeting the following additional requirements: 1. Real-time processing; 2. Multi-client support; 3. Minimized CPU utilization and PCI-E traffic; 4. Adaptive compression algorithm and ratio.

Contribution: Adapted a hardware-oriented, serialized algorithm to a massively parallel CUDA implementation on existing GPU. It meets requirements 1-3 and supports lossless compression. Algorithms are specially tuned to exploit the parallelism and memory hierarchy of NVIDIA GPUs. Novelty of the software solution includes redesigned program and data flow and hardware-specific optimizations. As of performance, the compressing module spent less than 4ms to process a frame of HD resolution (1920x1200) using GT200 (compared to over 130ms on an Intel E2180 CPU). For the first time, a single graphic card can handle HD frame compression for several to a dozen users at 30+fps.

Department of Computer Science, Cornell University, Ithaca, NY
Scaling BGP/MPLS VPN

Motive: In recent years, network-based VPN services have gained wide popularity. Consequently, routing and forwarding tables in provider equipment undergo a rapid growth. Observations show that the traditional any-to-any route provision is wasteful under existing traffic patterns. Scalability of network-based VPN services improves if routing and forwarding table size can be reduced significantly.

Goal: 1. Analyze topological and traffic characterizations of existing network-based VPN infrastructure. 2. Design a mechanism to reduce routing/forwarding table size without violating service-level agreements. 3. Implement the mechanism and integrate it into network infrastructure.

Contribution: Comprehensive analyses of collected data have been performed. We study basic aspects such as VPN deployment characteristics, infrastructure topology, traffic patterns and routing prefix utilization. Based on this insight, we find it plausible to introduce a tradeoff between worst-case path length and routing table size. An algorithm is designed to automatically configure route provisioning under performance constraints. We also suggest several simpler strategies to reduce routing table size.

Advisor: Professor Paul Francis

Converged Networks Group, IBM China Resarch Lab, Beijing, China
Internship in Performance Evaluation and Optimization of SIP-based Applications on IBM Java Virtual Machine

Motive: Despite their security advantage and development convenience, performance of servers running SIP-based applications on top of Java virtual machine (JVM) are often questioned, mainly due to the impact of GC. The basic GC policy in IBM JVM can be summarized as a stop-the-world mark-sweep-and-compact strategy. Consequently, GC introduces periodical slow response intervals and message loss.

Goal: 1. To streamline servers specifically for SIP-based applications by tuning a number of JVM policies, such as generational GC, concurrent marking and JVM parameters. 2. Calculate performance degradation curves of an optimized server configuration assuming certain user behavior, and point out the inflection point of performance.

Contribution: We deployed a series of experiments on our test bed with various JVM settings. So far, we are able to decide optimal policy configurations under different traffic patterns. A deterministic Petri Net model has also been developed to plot performance degradation curves and to predict the inflection point.

Department of Computer Science and Technology, Tsinghua University, China
Graduate Research Assistant in Enhance unstructured P2P system in Privacy and Security, a joint program with Microsoft Research Asia

Motive: While P2P systems have an edge in scalability, they are often quested about their fairness, privacy and security. On the other hand, efficiency, fairness and survivability are interwoven in many P2P file sharing systems such as BitTorrent. It is important to realize the interdependency of performance indices, and introduce enhancement into one or more aspects of such applications afterwards.

Goal: 1. To point out the how different aspects, such as fairness, survivability, efficiency and etc., are related to each other. 2. With a deep understanding of BitTorrent like networks acquired in preliminary research, put forward an improved design to enhance security and privacy in BitTorrent like file sharing systems. 3. Provide measurement tools to collect end-to-end performance metrics at low communication cost for a global view of P2P systems.

Contribution: Our preliminary research on takes a modeling approach on the dynamic behavior of BitTorrent like networks. The model reveals how fairness and survivability of a P2P file sharing system are related to the more traditional issues like scalability and efficiency. After realizing our first goal, an enhanced interconnecting mechanism has been designed and partially implemented to improve security and privacy, and this mechanism could be applied to general P2P file-sharing systems. Since achieving high security and privacy can often be contradictory, our scheme supports flexible adaptation when a compromise is required.

Advisor: Professor and Department Head Chuang Lin

Department of Computer Science and Technology, Tsinghua University, China

Graduate Research Assistant in Intel IXA University Program (www.ixaedu.com)

Motive: The functionality and performance of many network security applications rely on an ef-ficient, flexible implementation of cryptographic algorithms on network processors (NPs), especially on data plane (fast path). Necessity of benchmarking cryptographic algorithms on NPs rises in the face of rigid hardware solutions and unclear performance of software-based solutions.

Goal: 1. Implementing popular, representative cryptographic algorithms on network processors, with our focus on data-plane. 2. Benchmarking NPs to locate performance bottlenecks. 3. To alleviate the bottlenecks with generic or model-specific optimizations.

Contribution: A major developer of NPCryptBench, a cryptographic benchmark suite for network processors. By applying the benchmarks on several NP models we obtain first-hand statistics. With these data performance bottlenecks under different workloads and system configurations are successfully located. Summarized benchmark methodology and optimization guidelines from both software and architecture angles.

Advisor: Professor and Department Head Chuang Lin

State Key Laboratory of Intelligent Technology and Systems, Department of Computer Science and Technology, Tsinghua University, China

Undergraduate research in Decision Theory and Multi-agent Systems

This is a two semester discussion group/seminar whose members study the classical papers on decision theory and game theory used in multi-agent systems. At the end of each semester, I concluded the seminar with a survey essay. I also presented an introductory talk of multi-agent systems to my classmates and a short course in game theory to my graduate colleagues.

Advisor: Professor Mingsheng Ying and Assistant Professor Sanjiang Li