Modern networks typically perform a wide variety of functions, such as firewalling, intrusion detection, proxying, load balancing, NAT, or WAN optimization. Traditionally, these functions were implemented as middleboxes using dedicated hardware, but network function virtualization (NFV) has been moving this functionality to shared, cloud-like infrastructure. To remain transparent to the rest of the network, it is critical to ensure that the virtualized network functions can offer predictable latencies and guaranteed throughput, which existing clouds cannot provide. 

This project has developed theories and systems for building a scalable NFV infrastructure that can provide latency and throughput guarantees. To enable such guarantees, we have developed a range of novel techniques, including:

• Compositional analysis methods that efficiently analyze the timing behaviors of complex NFV applications using novel decomposition and resource-aware interfaces.
• Holistic real-time multi-resource allocation algorithms that provide strong isolation among concurrent applications while maximizing resource efficiency on modern multicore hardware.
• Efficient dynamic run-time adapation techniques for maximizing performance by exploiting the multi-phase execution of the application and the multi-mode nature of the system.
• Low-overhead kernel-bypass scheduling techniques that optimize tail latencies at microsecond scale.
• Novel detection algorithms and recovery protocols for handling crash faults and for defending against DDoS and Byzantine attacks.
• Methods for diagnosing timing performance problems using a novel time-aware extension of data provenance.
• Several implementation prototypes based on extensions of Xen, Linux, and LITMUS (an RTOS). 
• Interesting use cases beyond NFV, including data-center applications, cyber-physical systems, and robotics.

The project has advanced the state of the art not only in networking and cloud computing but also in other domains, such as security, real-time systems, cyber-physical systems, and robotics. The project has expanded the knowledge in a broad spectrum of research areas, including, e.g., scalable real-time scheduling and resource allocation algorithms for virtualized environments, data-driven dynamic run-time adapation techniques, low-overhead application-aware schedulers, novel timing diagnostic primitives, and new ways to defend against faults and attacks. The results enable the development of next-generation NFV platforms that will both simplify network management in data centers and support performance-critical applications. The results have also laid a foundation for time-predictable infrastructure that can support safety- and life-critical applications, such as self-driving vehicles and connected medical systems, as well as applications in IoT and robotics.

The project has helped to train eight PhD students, four of whom have graduated and accepted positions in the tech industry (e.g., Facebook and Apple) and two are close to graduation. It has also provided research experience and training to five Master's students and seven undergraduate students. Results from the project have been integrated with three courses at Penn, each of which has taught more than 100 undergraduate and graduate students per semester.

Last Modified: 01/25/2022
Modified by: Linh Thi Xuan Phan