Current Projects
Memory Systems and Long-Term RAM
Memory is now the largest share of what a server costs, and the cost per bit of DRAM has stopped falling. We research denser and cheaper classes of main memory, such as long-term RAM, and the operating system support they need to be usable. Rather than hide a new memory’s constraints behind a translation layer on the module, we are interested in what happens when a system exposes them and lets the kernel manage them: which data belongs in which memory, who may write it and when, and where the boundary between hardware and software should sit. Server memory is no longer one uniform resource, and how often a page is touched is too coarse a signal to place it well, so we also ask what hardware should tell an operating system about how well a page suits a given kind of memory. This work includes measurements of how long data lives and how it is accessed, kernel mechanisms for migrating and protecting pages, and studies of what disaggregated and compressed memory cost at datacenter scale.
Machine Learning Systems
Serving a large model is a systems problem before it is a modeling problem. What limits throughput is usually not arithmetic but the cost of moving data between devices. We research how to structure both models and the systems that run them so that communication stays local, inside a single device where that is possible and inside a single machine where it is not. This work includes distributed inference for models that route each input to a few of many experts, orchestration for applications assembled from several models and tools chained together, and measurement of where the time in a served model actually goes.
Verifiable and Private Computation
We research how to trust the result of a computation you did not run, and how to compute on data whose owners will not show it to you. Cryptographic proofs of computation have become practical in principle and remain awkward in practice, largely because the abstractions underneath them, memory above all, were not designed for the programs people want to prove things about. This work includes memory abstractions that make such proofs tractable at realistic scale, protocols that let the owner of a confidential model prove something about how it behaves without disclosing it, and coordination among parties who need a shared answer computed from data none of them will reveal.
Safe Systems Software in Rust
We research how far a type-safe language can be pushed as the foundation for systems software, and what it guarantees once it is. Much of this work goes into an embedded operating system written in Rust that runs several mutually distrustful applications on microcontrollers with no virtual memory and very little memory of any kind, and which has moved from a research prototype to production on millions of devices. The second question is the harder one. Systems built this way depend on language properties holding for code they did not write, and an extension that is safe by the language’s own definition can still break the invariants the system around it relies on. This work includes static analysis that finds those violations, case studies of where the gap appears in real systems, and the argument that it should be reasoned about directly rather than assumed away.
Past Projects
Distributed Systems for Computer Graphics
Computer graphics has yet to take advantage of the large-scale distributed systems of the past decade. With the exception of rendering, few graphics systems today are distributed. Instead, they run on powerful servers or, in the best case, a small cluster. Algorithms are designed for single hosts because distributing them is extremely hard. Unlike big data systems, which can logically partition data across keys, graphical systems typically have much more complex dependencies which are hard to distribute. However, the locality in these systems is not arbitrary, as it is typically bound to to a geometric representation of the problem. We are researching large-scale distributed systems that take advantage of this implicit locality to provide real-time, high performance graphics whose scale cannot be served by a few hosts. This research involves virtual world services such as queries and lookup, content simplification designed to load over a wide area network, and a runtime for distributed simulation.
This work is funded by the National Science Foundation (NeTS-ANET Grants #0831163 and #0831374) and conducted in conjunction with the Intel Science and Technology Center for Visual Computing.
Wireless Networking and Full Duplex Wireless
We research how to build robust, efficient wireless networks. Our work focuses on experimental and scientific results from real-world networks: how do they behave, why, and how can one leverage these behaviors to build better systems? This work includes detailed experimental studies of wireless networks, mathematical models to describe the behavior we observe, protocols which build on this deeper understanding, and novel radio designs for full duplex wireless that will enable a new generation of protocols.
This research has supported by generous gifts from DoCoMo Capital, the National Science Foundation under grants #0832820, #0831163, #0846014 and #0546630, the King Abdullah University of Science and Technology (KAUST), Microsoft Research, scholarships from the Samsung Scholarship Foundation, a Stanford Graduate Fellowship and a Stanford Terman Fellowship.
Embedded Sensing Systems
We research new programming languages, operating systems, and software for embedded devices, such as home automation networks, smart meters, and wireless sensors, as well as personal health devices such as Up and FitBit. Embedded systems are difficult computer platforms for programmers to work with. The need to meet hard real-time deadlines, fit programs into a small amount of memory, be robust to failures, and conserve energy wherever possible. Furthermore, ultra-low power systems have very different hardware, lacking abstractions such as virtualization and cache hierarchies. Our work focuses on software abstractions that allow you to build large, complex applications.
This work was supported by generous gifts from Intel Research, DoCoMo Capital, Foundation Capital, the National Science Foundation under grants #0615308 and #0846014, the King Abdullah University of Science and Technology (KAUST), Microsoft Research, scholarships from the Samsung Scholarship Foundation and a Stanford Terman Fellowship. Our scientific work was supported a National Science Foundation awards BCS-0947132, DE-AR-0000018, grants #0615308 and #0846014, a Branco Weiss fellowship, National Institute of Child Health and Human Development Award 1K01HD051494, and National Institutes of Health Grant GM28016.
Energy Efficient Computing
Energy is the limiting resource in a huge range of computing systems, from embedded sensors to mobile phones to data centers. We research how to design and build computer systems to manage energy and minimize its consumption. This work includes wireless sensing deployments to measure where energy goes in building-scale computer system, software techniques to track energy at microjoule accuracy on embedded systems, and operating systems for phones that make energy a first-class abstraction.
This work was supported by the Department of Energy ARPA-E program under award number DE-AR0000018, generous gifts from DoCoMo Capital, the National Science Foundation under grants #0832820, #0831163, #0846014 and #0546630, the King Abdullah University of Science and Technology (KAUST), Microsoft Research, scholarships from the Samsung Scholarship Foundation, a Stanford Graduate Fellowship and a Stanford Terman Fellowship.
Energy Systems and Datacenter Power
Energy is a first-class systems constraint, from grid-scale battery control to the power transients that large training clusters push back onto the grid.
Packet Processing and Memory Bandwidth
Network interfaces now have enough compute to inspect traffic at line rate and nowhere near enough memory bandwidth to feed it. We think the reason is an architectural mismatch: packets arrive variable-sized and out of order, while the memory they land in is a contiguous address space divided into fixed-size pages, and translating between the two layouts costs copies that consume exactly the bandwidth the analysis needed. We research addressing schemes and the hardware support that would let software read a whole flow as one contiguous region, and we build the datapaths on reconfigurable hardware to find out what they save. This work sits between networking, computer architecture, and memory systems, and none of the three can settle it alone.
Visibility as a Wireless Sensor Network Design Principle
We are pursuing a long-term agenda of research and education to improve the robustness, manageability, and scalability of low-power wireless sensing systems. The dominant design principle behind this agenda is network visibility. Described colloquially, visibility measures a user’s ability to identify the cause of a network event, such as a packet drop. We propose to research how to make networks more visible.
This project is funded by the National Science Foundation under award 0846014 (CAREER). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.