Reconfigurable Architectures for General-Purpose Computing

General-purpose computing devices allow us to (1) customize computation after fabrication and (2) conserve area by reusing expensive active circuitry for different functions in time. We define RP-space, a restricted domain of the general-purpose architectural space focussed on reconfigurable computing architectures. Two dominant features differentiate reconfigurable from special-purpose architectures and account for most of the area overhead associated with RP devices: (1) instructions which tell the device how to behave, and (2) flexible interconnect which supports task dependent dataflow between operations. We can characterize RP-space by the allocation and structure of these resources and compare the efficiencies of architectural points across broad application characteristics. Conventional FPGAs fall at one extreme end of this space and their efficiency ranges over two orders of magnitude across the space of application characteristics. Understanding RP-space and its consequences allows us to pick the best architecture for a task and to search for more robust design points in the space. Our DPGA, a fine- grained computing device which adds small, on-chip instruction memories to FPGAs is one such design point. For typical logic applications and finite- state machines, a DPGA can implement tasks in one-third the area of a traditional FPGA. TSFPGA, a variant of the DPGA which focuses on heavily time-switched interconnect, achieves circuit densities close to the DPGA, while reducing typical physical mapping times from hours to seconds. Rigid, fabrication-time organization of instruction resources significantly narrows the range of efficiency for conventional architectures. To avoid this performance brittleness, we developed MATRIX, the first architecture to defer the binding of instruction resources until run-time, allowing the application to organize resources according to its needs. Our focus MATRIX design point is based on an array of 8-bit ALU and register-file building blocks interconnected via a byte-wide network. With today's silicon, a single chip MATRIX array can deliver over 10 Gop/s (8-bit ops). On sample image processing tasks, we show that MATRIX yields 10-20x the computational density of conventional processors. Understanding the cost structure of RP-space helps us identify these intermediate architectural points and may provide useful insight more broadly in guiding our continual search for robust and efficient general-purpose computing structures.

Thread Scheduling Mechanisms for Multiple-Context Parallel Processors

Scheduling tasks to efficiently use the available processor resources is crucial to minimizing the runtime of applications on shared-memory parallel processors. One factor that contributes to poor processor utilization is the idle time caused by long latency operations, such as remote memory references or processor synchronization operations. One way of tolerating this latency is to use a processor with multiple hardware contexts that can rapidly switch to executing another thread of computation whenever a long latency operation occurs, thus increasing processor utilization by overlapping computation with communication. Although multiple contexts are effective for tolerating latency, this effectiveness can be limited by memory and network bandwidth, by cache interference effects among the multiple contexts, and by critical tasks sharing processor resources with less critical tasks. This thesis presents techniques that increase the effectiveness of multiple contexts by intelligently scheduling threads to make more efficient use of processor pipeline, bandwidth, and cache resources. This thesis proposes thread prioritization as a fundamental mechanism for directing the thread schedule on a multiple-context processor. A priority is assigned to each thread either statically or dynamically and is used by the thread scheduler to decide which threads to load in the contexts, and to decide which context to switch to on a context switch. We develop a multiple-context model that integrates both cache and network effects, and shows how thread prioritization can both maintain high processor utilization, and limit increases in critical path runtime caused by multithreading. The model also shows that in order to be effective in bandwidth limited applications, thread prioritization must be extended to prioritize memory requests. We show how simple hardware can prioritize the running of threads in the multiple contexts, and the issuing of requests to both the local memory and the network. Simulation experiments show how thread prioritization is used in a variety of applications. Thread prioritization can improve the performance of synchronization primitives by minimizing the number of processor cycles wasted in spinning and devoting more cycles to critical threads. Thread prioritization can be used in combination with other techniques to improve cache performance and minimize cache interference between different working sets in the cache. For applications that are critical path limited, thread prioritization can improve performance by allowing processor resources to be devoted preferentially to critical threads. These experimental results show that thread prioritization is a mechanism that can be used to implement a wide range of scheduling policies.

ADAM: A Decentralized Parallel Computer Architecture Featuring Fast Thread and Data Migration and a Uniform Hardware Abstraction

The furious pace of Moore's Law is driving computer architecture into a realm where the the speed of light is the dominant factor in system latencies. The number of clock cycles to span a chip are increasing, while the number of bits that can be accessed within a clock cycle is decreasing. Hence, it is becoming more difficult to hide latency. One alternative solution is to reduce latency by migrating threads and data, but the overhead of existing implementations has previously made migration an unserviceable solution so far. I present an architecture, implementation, and mechanisms that reduces the overhead of migration to the point where migration is a viable supplement to other latency hiding mechanisms, such as multithreading. The architecture is abstract, and presents programmers with a simple, uniform fine-grained multithreaded parallel programming model with implicit memory management. In other words, the spatial nature and implementation details (such as the number of processors) of a parallel machine are entirely hidden from the programmer. Compiler writers are encouraged to devise programming languages for the machine that guide a programmer to express their ideas in terms of objects, since objects exhibit an inherent physical locality of data and code. The machine implementation can then leverage this locality to automatically distribute data and threads across the physical machine by using a set of high performance migration mechanisms. An implementation of this architecture could migrate a null thread in 66 cycles -- over a factor of 1000 improvement over previous work. Performance also scales well; the time required to move a typical thread is only 4 to 5 times that of a null thread. Data migration performance is similar, and scales linearly with data block size. Since the performance of the migration mechanism is on par with that of an L2 cache, the implementation simulated in my work has no data caches and relies instead on multithreading and the migration mechanism to hide and reduce access latencies.

On-the-Fly Maintenance of Series-Parallel Relationships in Fork-Join Multithreaded Programs

A key capability of data-race detectors is to determine whether one thread executes logically in parallel with another or whether the threads must operate in series. This paper provides two algorithms, one serial and one parallel, to maintain series-parallel (SP) relationships "on the fly" for fork-join multithreaded programs. The serial SP-order algorithm runs in O(1) amortized time per operation. In contrast, the previously best algorithm requires a time per operation that is proportional to Tarjan’s functional inverse of Ackermann’s function. SP-order employs an order-maintenance data structure that allows us to implement a more efficient "English-Hebrew" labeling scheme than was used in earlier race detectors, which immediately yields an improved determinacy-race detector. In particular, any fork-join program running in T₁ time on a single processor can be checked on the fly for determinacy races in O(T₁) time. Corresponding improved bounds can also be obtained for more sophisticated data-race detectors, for example, those that use locks. By combining SP-order with Feng and Leiserson’s serial SP-bags algorithm, we obtain a parallel SP-maintenance algorithm, called SP-hybrid. Suppose that a fork-join program has n threads, T₁ work, and a critical-path length of T[subscript ∞]. When executed on P processors, we prove that SP-hybrid runs in O((T₁/P + PT[subscript ∞]) lg n) expected time. To understand this bound, consider that the original program obtains linear speed-up over a 1-processor execution when P = O(T₁/T[subscript ∞]). In contrast, SP-hybrid obtains linear speed-up when P = O(√T₁/T[subscript ∞]), but the work is increased by a factor of O(lg n).