Reservoir Labs Targets PERFECT Parallel Tools for DARPA

 
 
By Darryl K. Taft  |  Posted 2013-01-03
 
 
 

Reservoir Labs Targets PERFECT Parallel Tools for DARPA


Reservoir Labs announced it has been awarded a research contract with options up to $8.7 million in the Defense Advanced Research Projects Agency’s (DARPA) Power Efficiency Revolution for Embedded Computing Technology (PERFECT) program.

New York City-based Reservoir will be developing software to support new highly power efficient embedded computing devices. Resulting technology will support military systems as well as potentially lead to innovations in the mainstream IT and consumer spaces, as has been common over DARPA’s 50-plus year history.

Reservoir's goal in the PERFECT program is to help develop computer architectures that improve the power efficiency of embedded computing devices from today's 1 giga floating point operation per second per watt (GFLOPS/W) to a future 75 GFLOPS/W, the company said. This capability is needed for current and next generation military systems such as intelligence, surveillance and reconnaissance (ISR). Current computer architectures cannot meet the rapidly growing demands of ISR systems within the power, heat dissipation, and weight constraints of military platforms, and this gap is predicted to increase. This is because traditional Moore's Law scaling no longer can increase the performance of computers while maintaining the same power, the company said.

As DARPA's PERFECT program scales performance with new computing architectures, new programming tools and new algorithms will be required, agency officials said.

Reservoir's contribution to the program will be to advance the state of the art in programming tools for these new power efficient architectures. Reservoir will be leveraging its R-Stream optimization software for the effort. R-Stream uses operations research techniques and advanced polyhedral mathematics to generate efficient choreographies of computation over parallel computing hardware to avoid power-burning communications, said Richard Lethin, president of Reservoir Labs. Reservoir will be using R-Stream to target the new near-threshold computing processor architectures being developed in the PERFECT program and new algorithms for next generation ISR systems.

Reservoir Labs’ Software Power Optimization to Efficiency Revolution project, also known as SPOTTER, addresses the concurrency, locality, and algorithms elements in PERFECT. The goal of SPOTTER is to develop tools – compilers and runtimes -- and algorithms to enable a high-level, performance-portable programming model for 75 GFLOPS/W hardware, Lethin said.

"It is a great privilege to have been selected by DARPA for this research program,” Lethin said in a statement. “Our team is looking forward to the challenges presented by these new low power architectures and the opportunity to contribute to US security.  Furthermore, the technologies from the PERFECT program will have broad impact saving energy and enabling new capabilities for commercial and consumer devices as they rapidly adopt these innovations."

Lethin said a new class of multicore processors and accelerators has emerged with significant parallelism and on-chip distributed memory – such as General-purpose graphics processing units (GPGPU) -- that requires explicit control and communication among tiled computation kernels. This new class of computation devices offers over a tenfold improvement in power efficiency, but can be 100 times more difficult to program, he said.

The Reservoir technology will support the C language. The company also will be working with a domain-specific language known as Sensor Application Notation for Exascale or SANE, which was built with Air Force Research Labs Small Business Technology Transfer funding.

Reservoir Labs Targets PERFECT Parallel Tools for DARPA


Developing sensor applications such as image formation, change detection and adaptive imaging for the new architectures requires specialized skill. Developers must know about the app and its physics, required precision and algorithms, as well as the target hardware, parallelization and optimization techniques, Lethin said. This often involves an error-prone, two-step development process where first an application specialist employs a high-productivity language to create prototypes to validate the algorithm, and then a code jockey comes in to re-implement the app in a low-level language and parallelize the software. The objective of SANE is to lower the barrier and reduce requirements for success in writing code such that developers need only know about the app and not algorithms, target hardware and such.

Reservoir’s R-Stream is a high level compiler for embedded computing, parallel processing and HPC algorithms, and is designed to seamlessly generate parallelized code to target-specific or low-level C compilers, the company said. R-Stream is a source-to-source compiler that accepts a sequential C program as input and produces code that has been parallelized and optimized for the new types of processors.  R-Stream can output optimized code in a variety of formats for downstream processors, including highly optimized OpenMP and CUDA – formerly known as Compute Unified Device Architecture.

Moreover, R-Stream performs multiple advanced transformations from the input C source code to achieve high performance. These include special forms of array expansion -- to remove constraints on parallelism -- joint scheduling for parallelism and locality, task granularity selection, communications generation, software pipelining, memory region reshaping, and back end dialect generation. The resulting mapped program is more than simply parallelized — it represents a detailed choreography of computation and data motion across parallel units and through explicitly managed memory hierarchies, Lethin said.

Established in 1958, DARPA, then known simply as the Advanced Research Projects Agency (ARPA), is notorious for its involvement in the creation of the Internet, which began with an idea to link time-sharing computers into a national system. The agency’s ARPANET was a precursor to the Internet. A primary role of DARPA has been to spark innovations that can assist in the country’s war fighting capabilities. However, over the years many of the innovations that started at DARPA have found their way into the mainstream and into consumer and business uses.

Among the innovations that started at DARPA and moved into the mainstream are: Voice-to-text software; distributed computing; other projects that helped lead to innovations like Unix; Windows NT; packet switching; TCP/IP protocols; reduced instruction set computing; massively parallel processing; computer-aided design/computer-aided manufacturing; synchronous optical networking; asynchronous transfer mode and computer graphics.

Lethin said Reservoir hopes to add to that list. “The future of computing is ‘physical’ as embedded wireless sensors augmenting our reality,” he told eWEEK.  “The kinds of optimizations from our project, which are targeted at DOD ISR systems, will be useful for consumer devices performing video image processing, wireless communication, and other forms of sensing. It will also have application in big data processing.”

Lethin also noted that Reservoir will be releasing research versions of its compiler on a quarterly basis. The next release is Jan 15, 2013.

“We will license the compiler commercially for a fee and provide the compiler free to academics for research purposes,” Lethin said. “We will also provide evaluation versions. The compiler uses the Gurobi tool internally for optimization. That can be embedded in the commercially licensed versions and academics can obtain free licenses to Gurobi.”

Rocket Fuel