I am a graduate student working with Professors Randy Katz and Ion Stoica in the AMP Lab. Currently, my research is primarily focused on achieving predictable flow completion times in datacenter networks.
FastLane is a novel mechanism for reducing high-percentile flow completion times in datacenter networks. Traditionally, sources rely on indirect indicators of packet drops (3 duplicate ACKs or timeouts). These may take a long, highly variable amount of time to arrive. FastLane enlists switches, informing sources of packet drops as quickly as possible. Sources respond quickly and effectively, reducing high-percentile flow completion times by up to 75%, while avoiding the large buffer requirements of lossless interconnects.
DeTail proposes a network stack aimed at addressing three causes of unpredictable flow completion times: packet drops, uneven load balancing, and lack of prioritization. DeTail employs a lossless interconnect to prevent packet drops, per-packet load balancing to evenly spread load, and application-specified prioritization to ensure that latency-sensitive flows arrive on time. These mechanisms, combined with transport-level throttling of persistent congestion, reduce the high-percentile flow completion times that can stall datacenter workflows.
Incast is a complex phenomenon that can occur when multiple sources simultaneously send to the same destination. Flows can experience packet drops, timing out, and underutilizing the link. We propose simple, experimentally verified models that accurately predict when incast will occur and the throughput reduction it will cause. Then we evaluate how incast impacts Hadoop's performance. We show that while incast can significantly harm job completion time, performance overheads sometimes mask these gains.
Wireless sensor networks must support many different applications with different traffic workloads, while being as energy-efficient as possible. To address this issue, we implemented TSCH, a medium access control protocol that allows nodes to explicitly reserve timeslots in a TDMA schedule. By leveraging multiple channels, TSCH makes the network more resilient to interference and increases available throughput. We detail our implementation of TSCH and describe why this one protocol is well-suited to supporting event-based, periodic, and bulk traffic.
While localization is desirable in wireless sensor networks, it is difficult to achieve. Received signal strength (RSS) indicators suffer from constructive and destructive interference, while other techniques are too complex and power-hungry. We explore the advantages of using RF time-of-flight to obtain range estimates. Our prototype, Waldo, addresses timer limitations and mitigates multipath effects, demonstrating that a simple, low-cost system can achieve 1-3m accuracy.
Cisco Systems, Software Engineer, 06/2009 - 08/2009
Cisco Systems, Software Engineer, 06/2008 - 08/2008
Mainsail Communications, Software Engineer, 06/2007 - 03/2008
Graduate Student Instructor, 08/2008 - 12/2008 and 08/2009 - 12/2009
David Zats, Anand Iyer, Ganesh Ananthanarayanan, Randy Katz, Ion Stoica, and Amin Vahdat, “FastLane: Agile Drop Notification for Datacenter Networks”, Tech Report, October, 2013. [pdf]
David Zats, Tathagata Das, Prashanth Mohan, Dhruba Borthakur, and Randy Katz, “DeTail: reducing the flow completion time tail in datacenter networks”, ACM Sigcomm, Helsinki, Finland, August 2012. [pdf]
Yanpei Chen, Rean Griffith, David Zats, Anthony Joseph, and Randy Katz, “Understanding TCP Incast and Its Implications for Big Data Workloads”, USENIX ;login: Magazine, June 2012. [pdf]
Steven Lanzisera, David Zats, and Kristofer Pister, “Radio Frequency Time-of-Flight Distance Measurement for Low-Cost Wireless Sensor Localization”, IEEE Sensors Journal, March 2011. [pdf]
Ganesh Ananthanarayanan, David Zats, and Ion Stoica, “An AggreGATE Network Abstraction for Mobile Devices”, ACM Mobicom, Beijing, China, September 2009. [pdf]
Mohammad Rahimi, Shaun Ahmadian, David Zats, Rafael Laufer, and Deborah Estrin, “Magic of Numbers in Networks of Wireless Image Sensors”, Workshop on Distributed Smart Cameras, Boulder, Colorado, October 2006. [pdf]
Mohammad Rahimi, Shaun Ahmadian, David Zats, Rick Baer, Deborah Estrin, and Mani Srivastava, “Network of cyclops; image inference and interpretation in sensor network”, ACM SenSys, San Diego, California, November 2005. [pdf]