Small scale acoustic tomography for sensor networks (PhD thesis)
- Communications networks for underwater sensor networks have traditionally been designed to transport data and command/control information. In this work we re-envision the communications infrastructure of an underwater sensor network as a sensor itself. Inspired by ocean acoustic tomography, we observe that by precisely measuring the time-of-flight for communication signals sent in the network we can draw water temperature maps. To do so at small scales (inter-node distances on the order of 50 m to 1 km), requires very precise time-of-flight measurements. The noisy underwater channel also presents challenges. To date we have presented results from real-world experiments performed in Marina Del Rey that show water temperature measurements with sub-degree accuracy. Additionally, we have designed a coding technique that allows for combined data and sensing transmissions, thus reducing network overhead.
Pervasive Thermal Generation (CiSofT PTG)
- The PTG project is about increasing wireless sensor network deployments in the oil field. Previously, the deployment wireless sensor networks has been hindered by the cost of replacing batteries on industrial sensor nodes, like those from Emerson Rosemount. Many sources of latent energy in the form of heat are present in the oil field which represents a potential power source that can be tapped for use with wireless sensors. The delta-T (temperature difference) observed between the heat source and ambient was thought to be to low to allow for sufficient energy harvesting to power an industrial class wireless sensor node. With a novel (patent-pending) energy harvester and custom designed power conditioning circuit we were able to extract sufficient energy from an oil field flow line to indefinitely power a Rosemount pressure sensor.
In-situ digital power monitoring
- Providing accurate, low-latency on chip power monitoring for complex digital circuits like CPUs or GPUs is a sought after goal that would enable a new era in energy efficient computation. For example, knowing the power usage (and thereby temperature) of one core of a multi-core CPU could allow the OS scheduler to better manage the total CPU power by scheduling lower intensity jobs on a hot-running core. However, due to limitations with measurement techniques and power modeling, such power monitoring has to date not been possible. However, through a combination of activity counters, external total power measurements and a numerical algorithm, we have created a real-time, extremely accurate, sub-component power monitoring system. Additionally, we have shown similar techniques can be used to derive energy-per-instruction for a specific CPU, thus allowing accurate power measurements to be made using hardware instruction profiling, or accurate power estimates to be made using offline code profiling. Further, since the technique is built into the chip, it can be used during pre-packaging testing to increase yield.
Software defined networking for the Smart Grid
- In an effort to increase wide area situational awareness regarding the national power grid, building a "Smart Grid" has become a major collaborative effort at many institutions. One aspect of this work deals with the delivery of real-time monitoring data about the health of the power grid. The data is collected by synchronous phasor measurement units (PMUs) and must be delivered across a computer network with high-reliability and very low latency. The network semantics is that of a "publish-subscribe" network model, and thus are amenable to implementation with software defined networks (SDN). In this work we show how SDN can be used to implement the required semantics of a PMU network in a manner superior to current and proposed solutions.
Distributed DNA read mapping
- Increasingly bio-informatic applications such as DNA sequence mapping are seen as "big-data" applications, however it is not always straight forward to adapt current algorithms into a distributed algorithm that can be run on cloud-like distributed computing platforms. In this work we started with a state-of-the art DNA read mapping algorithm (PerM) developed in the bio-informatics department at USC. The algorithm is clever, however the existing implementation was inefficient and not distributable. We identified three ways to implement the same algorithm in a distributed fashion and using a cloud computing platform at ISI we examined the performance of the three methods while increasing both the number of nodes available and the size of the dataset. With these results we are able to choose the best method to use given the number of nodes available and dataset size. Additionally, we implemented a version of the PerM algorithm using the Map-Reduce distributed processing paradigm, which results in a very simple, elegant implementation that is easy to understand and modify.