Cornell Systems Lunch

In this talk I will present several works that concern with offloading.

I will first present NICOS - a Network Interface Card OS that enables a developer to dynamically offload custom tasks to the NIC. This work also provides an innovative scheduling scheme for non-preemptive environments. Next, I will review the state-of-the-art in processor technology and programmable peripheral devices. Finally, I will present Hydra - a generic code offloading framework. Hydra enables an application developer to design the offloading aspects of the application by specifying an "offloading layout", which is enforced by the runtime, during application deployment. The framework also provides the necessary development tools and programming constructs for developing such applications.

Because of the topic's relevance, non-cooperation in peer-to-peer computing has become an active field of research. In the first part of this talk, I will present BitThief, a free-riding BitTorrent client which applies simple tricks to avoid uploading while still maintaining good download rates. In the second part of my talk, a game-theoretic model is introduced which helps to gain insights into the effects of non-cooperation on a given distributed system.

Mobile computing systems rely on radio frequency communications to support interactions with larger distributed systems such as the Internet and the World Wide Web. While logically convenient, the view of RF communications as a point-to-point link is unrealistic, particularly in RF-dense urban environments. The RF communication channel is affected by many factors in the urban setting, including shadowing and multipath.

Distributed Radio combines multiple radio devices into a system, deriving benefits from the cooperative sharing of resources. While the processing resource sharing is not unlike that of a distributed computing system, the treatment of location as a resource enables a potentially much more reliable system. A useful view is that the positions of the elements in a distributed radio provide a statistical sample of possible locations, from which the one(s) most suitable for a particular task (e.g., receiving a GPS signal) can be selected.

The statistical sampling perspective leads to an extension of the distributed radio to mobile nodes. We introduce the idea of "mobility gain" and show how it can be obtained by self-positioning radio packet repeaters.

This talk discusses COPE, an alternate architecture for wireless mesh networks. In addition to forwarding packets, routers mix (i.e., code) packets from different sources to increase the information content of each transmission. I will show that intelligently mixing packets increases network throughput. COPE's design is rooted in the theory of network coding. Prior work on network coding is mainly theoretical and focuses on multicast traffic. This talk bridges theory with practice. It addresses the common case of unicast traffic, dynamic and potentially bursty flows, and practical issues facing the integration of network coding in the current network stack. We evaluate COPE on a 20-node wireless network, and discuss the results of the first testbed deployment of wireless network coding. The results show that COPE largely increases network throughput. The gains vary from a few percent to several folds depending on the traffic pattern, congestion level, and transport protocol.

Dina Katabi is an Associate Professor in the Electrical Engineering and Computer Science Department at MIT. She has joined the MIT faculty in March 2003, after completing her PhD at MIT. Dina's work focuses on wireless networks, network security, routing, and distributed resource management. She has award-winning papers in ACM SIGCOMM and Usenix NSDI. Further, she has been awarded a Sloan Fellowship award in 2006, the NBX Career Development chair in 2006, and an NSF CAREER award in 2005. Her doctoral dissertation won an ACM Honorable Mention award and a Sprowls award for academic excellence.

Kurzweil says, computers will enable people to live forever and doctors will be doing backup of your memories by late 2030. This talk is not about that, yet. Instead, the remarkable drop in disk costs makes it possible and attractive to retain past application states and store them for a long time for mining or auditing. A still open question is how to best organize the past state storage? Split snapshots are a recent approach to past state storage that is attractive for several reasons. Split snapshots are persistent, can be taken with high-frequency, and they are transactionally consistent. Unmodified database code can run against them. Like no other past state storage approach, they provide low-cost discriminated garbage collection of snapshots, a useful capability in long-lived systems since since indiscriminately keeping all snapshots accessible is impractical even if raw disk storage is cheap, because administering such large-volume storage is expansive over long duration.

A number of novel techniques underly split snapshots. A new in-memory data-structure creates consistent copy-on-write snapshots without blocking, a new persistent data structure provides high performance versioned meta-data, and a new snapshot storage organization allows to gradually garbage collect selected copy-on-write snapshots without creating disk-fragmentation and without copying. Measurements of a split snapshot prototype system indicate that the new techniques are efficient and scalable, imposing minimal (4%) performance penalty on a storage system, on expected common workloads.