Introduction
In 1986, researchers at Carnegie Mellon University introduced Agora, a programming environment designed to address the "productivity crisis" in developing complex, parallel AI systems such as speech and vision recognition. By providing a virtual machine layer that abstracts hardware and language differences, Agora allows researchers to focus on algorithmic innovation rather than the low-level "plumbing" of parallel processing.
The Productivity Crisis in AI
During the mid-1980s, high-performance AI applications faced a significant bottleneck. Systems like speech recognition were computationally intensive, requiring more power than a single processor could provide, yet they were notoriously difficult to parallelize. These systems are characterized by:
- Heterogeneity: They often require multiple languages (e.g., C for signal processing, Lisp for logic) and diverse hardware architectures.
- Rapid Evolution: Constant algorithmic changes make hard-coded parallel implementations brittle and difficult to maintain.
- Loose Structure: Unpredictable data flows make standard vectorization or rigid parallel models ineffective.
Before Agora, assembling a parallel version of a speech system could take over six man-months of grueling low-level programming.
The Agora Architecture
The authors proposed a layered architecture to decouple application logic from underlying hardware, as illustrated in the system's structural overview.
- The Mach Layer: Built on the Mach operating system, providing basic abstractions like message passing, shared memory, and threads.
- The Parallel Virtual Machine (Agora): The core "assembly language" of the system. It utilizes Knowledge Sources (KS)—independent procedures written in C or Lisp—and Cliques, which are global, named data structures. KSs are activated via pattern matching; when specific data arrives in a clique, the corresponding KS triggers.
- Frameworks: High-level, domain-specific "mini-environments" (e.g., data-flow editors) that allow researchers to define systems using familiar terms while the framework handles mapping to the virtual machine.
- Custom Hardware Integration: To overcome general-purpose OS overhead (which limits task granularity to ~10ms), custom VLSI architectures were designed for search algorithms, enabling much finer-grained parallelism.
Evaluation: The ANGEL Speech System
Agora was evaluated using the ANGEL speech recognition system. The system utilized real-world speech data to test "best match, beam search" algorithms across vocabularies ranging from 200 to 5,000 words.
A key insight from this evaluation was that productivity is a vital metric. Agora reduced the time required to assemble a complex parallel system from six months to just one month. Furthermore, the researchers identified a "granularity bottleneck": general-purpose CPUs are often inefficient for the micro-tasks required in AI search trees, necessitating custom hardware.
Results and Performance
The performance results demonstrated significant scaling capabilities. As shown in the scaling analysis, the speedup over a VAX-11/780 increases dramatically with the addition of processors and memory blocks.
Key results include:
- Development Speed: System assembly time was cut by over 80%.
- Performance Scaling: A configuration with 5 Knowledge Sources and 1 memory block achieved a 170x speedup over a VAX-11/780. A larger configuration with 28 Knowledge Sources and 7 memory blocks achieved a 1000x (three orders of magnitude) speedup, theoretically enabling real-time recognition of a 5,000-word vocabulary.
Conclusion
Agora demonstrated that a heterogeneous environment could significantly reduce the complexity of building large-scale parallel AI applications. By abstracting hardware and language complexities, it allowed researchers to achieve massive performance gains while maintaining high development productivity. Future work aimed to automate the clustering of Knowledge Sources and further reduce the "assembly-to-run" time to a single day.
Summary updated