Projects and Experiences

Randolph Y. Wang

I have done some very different things over the years and have at times taken somewhat less conventional paths. In this document, I briefly describe some of my past projects. In the future, I see myself continuing to be open to possibly pursuing new and different things as well as more traditional paths.

Intel (2010-2015)

The Edison team at Intel Labs China consists of people working closely at hardware (including SoC), operating system, middleware, application and device levels. My role as the lead includes conception of the overall project, driving key aspects of the architecture (such as the loosely coupled duo-core architecture, and the file system modifications to support streaming), coordinating design features that require cross-layer collaboration (such as low-power mode operations), and conceiving and designing the demonstration device and application details.

Hardware basics

Intel Edison: a tiny computer exactly the size and shape of an SD card.
Plug-able and unplug-able. For easily building certain types of IoT devices and applications.
Includes Wifi and Bluetooth, flash storage, DDR, and some sensors.
Custom-designed SoC. Two largely independent cores. Big core runs a variation of Android. Small core runs a thin layer of RTOS and/or Arduino. The cores share I/Os.
(The Intel Labs version of Edison, a research prototype, won numerous CES awards in 2014. The released product version is unfortunately bungled in my personal opinion and bears little resemblance to our original prototype and has not been well received.)
Reason for having two cores: big core gives rich functionality and the ability to “do apps.” The small core can put the big core to sleep, wake it up when necessary, and keep less demanding functionalities running, all the while conserving power.
The whole system has a peak power consumption under 1w (and can be powered by a standard SD card slot alone). (The 1w peak power happens when both cores are running at peak speeds, and Wifi, DDR, flash are all kept busy.) Lower power modes in 10s of milliwatts.
In 2014, the BOM of the whole thing was under $10.
Roughly in the space of existing boards such as Raspberry Pi and Arduino. One key strong point we provide is a unique combination of powerful capabilities coupled with extreme small form factor and low power. Size matters (so does power): it could enable new devices and scenarios not possible before.
We’re currently designing a next-generation version of Edison with new SoC features.

What it’s for

For quickly and easily building a certain class of “consumer IoT” devices.
(Caveat. “IoT” is a big and over-used word: there are many use cases and scenarios that we do not claim the machine to be suitable for.)
Advantages (already mentioned): small, low-power, powerful capabilities for supporting apps, Arduino-ready MCU, plug-able, unplug-able, upgradeable, good support for device collaboration via the cloud, easy to use or code for, …
Many demo devices had been built by us over the years to demonstrate the unique advantages.
The most polished one is a porcelain mug with touch, an LED matrix, and other assorted Edison-powered electronics embedded inside the thin porcelain wall. It can download and run arbitrary “mug apps.” (Dishwasher-safe, can withstand 100 degree hot water…)

Digital StudyHall (2005- )

In 2005, I co-founded a non-profit organization named the Digital StudyHall, an organization mainly working in India and a few other South Asian countries. The aim is to help improve children’s education in some of the most impoverished rural areas and urban slums. The environment in these places is such that we can only rely on the simplest low-cost technology solutions.

What we do is the following. We recruit top teachers in urban schools catering to middle-class children. We film the live lessons based on the local government school curriculum, in local languages. The content is comprehensive: almost on a page-by-page basis of the local curriculum school books. To ensure appropriateness of the filmed lessons for their eventual audience in much poorer areas, the top teachers are filmed teaching children from neighborhood slums: this ensures that children of similar disadvantaged backgrounds would find the filmed lessons to contain content, language, and methodology that are familiar to them. These lessons are organized at grassroots level in a decentralized manner in many cities by affiliated volunteer sub-organizations, in the local languages of the individual urban centers, an approach that allows us to scale.

In 2005, these lessons were filmed on cheap consumer grade tape camcorders, with a lapel microphone worn by the filmed teachers. The equipment cost is very low. The tapes are funneled to urban “digitization centers” where the tapes are digitized, slightly edited, stored in decentralized databases, and annotated with metadata. In some cases, supplementary information such as quizzes, subtitles, and exercises are also added as companion materials to the database.

From these urban centers, the resulting lessons are then distributed to impoverished rural and slum areas that share the same local language and government-stipulated syllabus. These areas have no Internet, computer, or even electricity so we must use the simplest available but practical technology. In 2005, the lessons were disseminated on DVDs. We help the urban digitization centers set up systems to generate these DVDs on a large scale. The recipient schools were given donated TVs and DVD players. In some other cases, we also used small projectors. Depending on the environment, we also help the schools devise electricity-generating solutions, including solar and diesel generators. Again, the equipment cost is kept low.

In the recipient schools, it is important to recognize that making children simply “watch TV” alone is not effective. Instead, we require local “mediator teachers” to play and pause the videos and provide interaction with the local children. These interactions may include board work, questions and answers, games, and other activities that are designed to keep the children engaged. So, in effect, the local children are being taught by two teachers at the same time: one “virtual teacher” on TV who understands the material and pedagogy well, one local mediator teacher who is less skilled but provides the crucial human interaction element. What we have consistently found over the years is that it’s this combination that makes the system work. Over time, the local mediator teachers may improve along the TV teacher so this also serves as a powerful teacher training tool.

Over the years, in conjunction with education experts from US universities, we have conducted vigorous comparative studies, published in education journals. Making substantial improvements in the local schools, in itself, is not sufficient evidence of the effectiveness of our approach; this is because the performance of these schools prior to our intervention was so poor that improving upon this typically abysmal baseline hardly constitutes evidence of effectiveness of our system. Instead, what we have systematically evaluated is how well the students that have been taught by the combination of the TV teachers and the local mediators performed in comparison with how well the urban center students that have been taught by the live top teachers performed; in our opinion, this is a high bar to pass; the results were very close. Over the years, the Digital StudyHall has won numerous awards. One key thing that we have learned is how to make the simple technology-based elements and the people-centric elements (such as pedagogy and organization) work seamlessly together to accomplish a high-level goal.

Princeton (1999-2006)

I was an assistant professor at the Computer Science Department at Princeton University. I conducted research in various systems areas including distributed storage systems, mobile storage systems, reliability and security of wide-area routing protocols, high-performance disk arrays, high-performance local file systems, and power consumption of storage technologies.

Reliable and secure wide-area routing protocols. The series of papers have the common theme of monitoring and exploiting multiple paths across the wide area to increase reliability and security of the high-level client systems. Two specific examples are the following. First, wide-area network services running across geographically-dispersed end points allow us to collect more complete and more fine-grained global information more efficiently than simpler approaches. Second, while taking advantage of multiple paths might seem like an obvious idea, one difficulty that we address is over-aggressiveness in cases where the multiple paths being exploited might encounter shared bottlenecks experiencing congestion.

Distributed storage systems. We investigate approaches that (1) distribute and locate data in a mobile environment so that as elements of the network move around and as the network quality is non-uniform, a user continues to enjoy the best possible availability and performance; (2) embed storage elements inside the network infrastructure and make these storage nodes both programmable and aware of network topologies to deliver the best possible availability and performance; (3) construct load-balanced building blocks of a networked file system that can not only mask less serious failures but also isolate more severe failures as the amount of failure increases in the system but as many clients as possible still enjoy uninterrupted access to parts of system services.

High-performance disk arrays and local file systems. One of the key ideas of these systems is that by precisely tracking and predicting the location of the rapidly moving disk heads in real time with respect to on-disk data structures, and by judiciously placing replicas at strategic locations on disks, one can significantly improve the performance of disk arrays. We also demonstrate how the well-known log-structured approaches can be applied in innovative ways to both a disk head location-aware local file system and a mobile storage system encompassing multiple logs to provide better performance and reliability than traditional systems can.

Power consumption of mobile storage systems. We have developed a highly accurate power consumption simulator of hard disks. We have investigated how different types of storage technologies perform in terms of their power consumption characteristics when coupled with different power management mechanisms, file system alternatives and policies, and different workload characteristics.

Berkeley (1991-1998)

I was a grad student at the computer science department at UC-Berkeley. Most of my publications from that period had to do with distributed network file systems, in terms of how to distribute file system responsibilities and data across a large number of servers to make the whole system more scalable. A number of other publications had to do with modeling and optimizing performance of high performance networks employed in a number of supercomputers. There was also an early paper on how to dynamically adjust lossy compression rates of a packet video system to adapt to fluctuating available bandwidth and packet losses of a wide-area network.