Shobana Padmanabhan

Extracting and improving microarchitecture performance on reconfigurable architectures

Applications for constrained embedded systems require careful attention to the match between the application and the support offered by an architecture, at the ISA and microarchitecture levels. Generic processors, such as ARM and Power PC, are inexpensive, but with respect to a given application, they often overprovision in areas that are unimportant for the application’s performance. Moreover, while application-specific, customized logic could dramatically improve the performance of an application, that approach is typically too expensive to justify its cost for most applications. In this paper, we describe our experience using reconfigurable architectures to develop an understanding of an application’s performance and to enhance its performance with respect to customized, constrained logic. We begin with a standard ISA currently in use for embedded systems. We modify its core to measure performance characteristics, obtaining a system that provides cycle-accurate timings and presents results in the style of gprof, but with absolutely no software overhead. We then provide cache-behavior statistics that are typically unavailable in a generic processor. In contrast with simulation, our approach executes the program at full speed and delivers statistics based on the actual behavior of the cache subsystem. Finally, in response to the performance profile developed on our platform, we evaluate various uses of the FPGA-realized instruction and data caches in terms of the application’s performance.

Optimal design-space exploration of streaming applications

Many embedded and scientific applications are pipelined (i.e., streaming) and deployed on application-specific systems. Typically, there are several design parameters in the algorithms and architectures used that impact the tradeoff between different metrics of application performance as well as resource utilization. Efficient automatic exploration of this design space is the goal of our research. We present a global optimization framework comprising a domain-specific variation of branch-and-bound that reduces search complexity by exploiting the topology of the applications pipelining. We exploit the topological information to discover decomposability through the canonical Jordan block form. The reduction in search complexity for four real-world streaming applications (drawn from the literature) is significant, ranging from a million-fold reduction in search space size to a reduction factor of 10 billion. All four optimization problems are thereby solvable in reasonable time.

Automatic application-specific microarchitecture reconfiguration

Applications for constrained embedded systems are subject to strict time constraints and restrictive resource utilization. With soft core processors, application developers can customize the processor for their application, constrained by resources but aimed at high application performance. With such freedom in the design space of the processor, however, comes complexity. We present here an automatic optimization technique that helps the developers with the processor microarchitecture customization. A naive approach exploring all possible configurations is exponential with the number of parameters and hence is clearly infeasible, even with only tens of reconfigurable parameters. Instead, our approach runs in time that is linear with the number of parameter values, based on an assumption of parameter independence. This makes the approach feasible and scalable. For the dimensions that we customize, namely application runtime and hardware resources, we formulate their costs as a constrained binary integer nonlinear optimization program. Though the results are not guaranteed to be optimal, we find they are near-optimal in practice. Our technique itself is general and can be applied to other design-space exploration problems

Unchaining in Design-Space Optimization of Streaming Applications

Data-streaming applications are frequently pipelined and deployed on hybrid systems to meet performance requirements and resource constraints. With freedom in the design of algorithms and architectures, the search complexity can explode. A popular approach to reducing search complexity is to decompose the search space while preserving optimality. We present a novel decomposition technique called unchaining that partitions the problem such that the resulting sub problems are less complex. Thanks to unchaining, the number of sub problems from the decomposition is linear in the number of chained blocks in the variable-constraint matrix (instead of being their product). Finally, we present a queueing network model and the quantitative search space reduction for a real world implementation of a bio sequence search application called BLASTN.

Convexity in Non-convex Optimizations of Streaming Applications

Streaming data applications are frequently pipelined and deployed on application-specific systems to meet performance requirements and resource constraints. Typically, there are several design parameters in the algorithms and architectures used that impact the application performance as well as resource utilization. Efficient exploration of this design space is the goal of this research. When using architecturally diverse systems to accelerate streaming applications, the design search space is often complex. The search complexity can be reduced by recognizing and exploiting convex variables to perform convex decomposition, preserving optimality even in the context of a non-convex optimization problem. This paper presents a formal treatment of convex variables and convex decomposition, including a proof that the technique preserves optimality. It also quantifies the reduction in the search space that is realized, at minimum equal to the number of distinct values of the convex variable and potentially much higher.

Decomposition techniques for optimal design-space exploration of streaming applications

Streaming data programs are an important class of applications, for which queueing network models are frequently available. While the design space can be large, decomposition techniques can be effective at design space reduction. We introduce two decomposition techniques called convex decomposition and unchaining and present implications for a biosequence search application.

Performance Tuning of Streaming Applications via Search-space Decomposition

High-performance streaming applications are typically pipelined and deployed on architecturally diverse (hybrid)systems. Developers of such applications are interested in customizing components used, so as to benefit application performance. We present an efficient and automatic technique for design-space exploration of applications in this problem domain. We solve performance tuning as an optimization problem by formulating cost functions using results from queueing theory. This results in a mixed-integer nonlinear optimization problem which is NP-hard. We reduce the search complexity by decomposing the search space. We have developed a domain-specific decomposition technique using topological information of the application embodied in the queueing network models. Our analysis includes when our decomposition preserves optimality. Our preliminary empirical results confirm two-fold benefits--solving a problem that is currently not solvable using state-of-the-art solvers and in some problem instances, improving initial solution value from the solver by over two orders of magnitude. Type of Report: Other Department of Computer Science & Engineering Washington University in St. Louis Campus Box 1045 St. Louis, MO 63130 ph: (314) 935-6160 Performance Tuning of Streaming Applications via Search-space Decomposition Shobana Padmanabhan, Roger D. Chamberlain, and Yixin Chen Dept. of Computer Science and Engineering, Washington University in St. Louis

Hierarchical Clustering using Reconfigurable Devices

Non-hierarchical k-means algorithms have been implemented in hardware, most frequently for image clustering. Here, we focus on hierarchical clustering of text documents based on document similarity. To our knowledge, this is the first work to present a hierarchical clustering algorithm designed for hardware implementation and ours is the first hardware-accelerated implementation