Glenn Reinman - US grants

Speculative execution has been widely used in computer architecture
research in the pursuit of higher levels of instruction level
parallelism. However, recent studies have shown that this technique
can significantly impact the scalability of the processor pipeline.
This research will investigate a number of speculative techniques
with respect to performance, scalability, energy efficiency, area
costs, and complexity; eventually leading to a novel, scalable
architecture to provide high performance at future technology sizes.
Such an architecture will minimize the amount of logic on the
critical timing path of the processor, relying on small speculative
structures to reduce the cycle time of the processor. Off the
critical path, a variety of larger speculative structures provide a
backing store to the critical path.

Some of the techniques to be examined in this research include
branch prediction, value and address prediction, data prefetching,
early register release and late register allocation, and multiple
clustered functional units. Many of these techniques require large
and often complex predictors that can impact the cycle time and power
density of a processor. This research investigates some techniques
to reduce the access time and size of these structures, while still
providing the accuracy to avoid costly speculative recovery
procedures.

Because of the relentless march of the silicon-based electronics technology as predicted by Moore's Law, computation, storage, and communication are now woven into the fabrics of our lives. The emerging technology of flexible electronics, where electronics components such as transistors and wires are built on a thin flexible material, offers a similar opportunity to weave computation, storage, and communication into the fabric of the very clothing that we wear. The implications of seamlessly integrating a large number of communicating computation and storage resources, mated with sensors and actuators, in close proximity to the human body will transform many aspects of biomedical research and practice. For example, one can imagine biomedical applications where biometric and ambient sensors are woven into the garment of a patient or a person in a medically-critical or hazardous environment to trigger or modulate the delivery of a drug.
To realize this vision outside the laboratory, radical innovation is required in the area of system-level information technology. These systems will not scale to widespread use if they are viewed simply as traditional chips or motherboards based on a different, flexible form factor. Rather, a rethinking of the architecture and the design methodology for all layers of these systems is needed. The reasons are two-fold. First, the underlying technology of electronics in flexible materials has characteristics and computation-communication cost trade-offs that are very different from that of silicon and PCB-based electronics. Second, the natural applications of these systems have environmental dynamics, physical coupling, resource constraints, infrastructure support, and robustness requirements that are very different from those faced by traditional systems.
One of the challenges in developing the needed information technology architecture and design methodology for these systems is that one needs to both conduct experimental work and develop a conceptual understanding of the problem domain. This research studies:
Application: Use as a driver application capability, reconfigurable fabric (R-Fabric) based on a combination of (i) the technology of flexible electronics using organic materials, and (ii) computing, communication, and sensing elements implemented as E-Buttons.
Architecture: Develop the general architecture concepts and cost/performance optimization techniques. The issues that we will focus on will include (i) appropriate primitives for composing the architecture, (ii) system interconnect network optimized for the electrical characteristics of the organic electronics, (iii) techniques to cope with the high ration of communication to computation cost, and (iv) architecture level self-configuration and re-configuration for robust operation.
Programming: Develop techniques and primitives for programming a system composed of hundreds of computation, storage, sensing, and actuation elements that are individually resource constrained and are connected by a structured but fault-prone high-cost interconnect network.
Processors: Develop domain-specific processor architecture optimized for these power-constrained, physically coupled applications.

Design Methodology: Develop techniques and hybrid emulation platform for systematic architecture exploration, simulation, optimization, and reconfiguration of these systems.

"This award is funded under the American Recovery and Reinvestment Act of 2009
(Public Law 111-5)."


Customizable Domain-Specific Computing
To meet ever-increasing computing needs and overcome power density limitations, the computing industry has entered the era of parallelization, with tens to hundreds of computing cores integrated into a single processor; and hundreds to thousands of computing servers connected in warehouse-scale data centers. However, such highly parallel, general-purpose computing systems still face serious challenges in terms of performance, energy, heat dissipation, space, and cost. In this project we look beyond parallelization and focus on domain-specific customization as the next disruptive technology to bring orders-of-magnitude power-performance efficiency improvement to important application domains.
The intellectual merit of this project includes development of a general methodology for creating novel customizable architecture platforms and the associated compilation tools and runtime management environment to support domain-specific computing to: 1) achieve orders-of-magnitude computing efficiency improvement for applications in a specific domain; and 2) demon-strate that such improvement can be obtained with little or no impact on design productivity, so that it can be deployed in a wide range of application domains. Our proposed domain-specific customizable computing platform includes: 1) a wide range of customizable computing elements, from heterogeneous fixed cores to coarse-grain customizable cores, and to fine-grain field-programmable circuit fabrics; 2) customizable high-performance radio frequency interconnects; 3) highly automated compilation tools and runtime management software systems for application development; and 4) a general, reusable methodology for customizable computing applicable across different domains. By combining these critical capabilities, we shall deliver a super-computer-in-a-box that is customized to a particular application domain to enable disruptive innovations in that domain. This approach will be demonstrated in several important application domains in healthcare.
The broader impact of this project will be measured by the new digital revolution enabled by customized computing. We will demonstrate the feasibility and advantages of the proposed research in the domain of healthcare, given its significant impact on the national economy and quality of life issues. In particular, we focus our effort on revolutionizing the role of medical imaging and hemodynamic modeling in healthcare, providing much more cost-efficient, convenient solutions for preventative, diagnostic, and therapeutic procedures to dramatically improve healthcare quality, efficiency, and patient outcomes. The broader impact of this project also includes the integration of research and education, exposing graduate, undergraduate, and high school stu-dents to the new concepts and research from this project via several new courses jointly developed and shared by researchers in our newly established Center for Domain-Specific Computing (CSDC). Summer research fellowship programs to support high school and undergraduate students will be provided by CSDC. Our goal is to train a new generation of students who are prepared for customized parallelization and computing, and can effectively apply such techniques to many areas of our society, thus furthering the digital revolution. Special efforts are being made to attract underrepresented students at all levels via partnerships with campus organizations focused on diversity, such as the UCLA Center for Excellence in Engineering and Diversity.
This research will be carried out as a collaborative effort between four universities: UCLA (the lead institution), Rice, UC Santa Barbara, and Ohio State. The research team consists of a group of highly accomplished researchers with diversified backgrounds, including computer science and engineering, electrical engineering, medicine, and applied mathematics. For more information, please visit http://cdsc.cs.ucla.edu.

Many healthcare applications present significant computational challenges. For example, the computational demand for personalized cancer treatment is prohibitively high for the general-purpose computing technologies, as tumor heterogeneity requires great sequencing depths, structural aberrations are difficult to detect with today's methods, and the tumor has the ability to evolve i.e., the same tumor might be assayed a great many times during the course of treatment. The goal of this project is to apply the domain-specific customized computing techniques developed by the Center for Domain-Specific Computing (CDSC) at UCLA to greatly accelerate computation for some key healthcare applications.

The CDSC, established in 2009 with the support of the NSF, looks beyond parallelization, and focuses on domain-specific customization as the next disruptive technology for power-performance efficiency improvement. In the past four years, CDSC has demonstrated significant performance and energy efficiency with innovation in developing customizable heterogeneous computing technologies. The current proposal under the NSF Innovation Transition program leverages the research results from CDSC, and focuses on key research problems and solutions to make domain-specific customizable computing feasible and practical for innovation transition to the industry, Specifically, the project will develop accelerator-rich architectures along with unified adaptive runtime systems for personalized cancer treatment, medical image processing, and will enable deployment in several energy efficient programmable platforms capable of handling huge volumes of state of the art real time patient data.

The center will continue its already successful outreach program, through a partnership with the UCLA Center for Excellence in Engineering and Diversity, to involve highly diversified high school and undergraduate students for summer research. The success of our project will enable significant advances in medical imaging analysis and personalized cancer treatment, which will greatly improve healthcare quality while reducing cost. The participation of the industrial partner in this InTrans project will greatly facilitate the innovation transition of research results from this project to industry for energy-efficient computing.