Morphable Cache Architectures: Potential Benefits

  • Authors:

  • I. Kadayif;M. Kandemir;N. Vijaykrishnan;M. J. Irwin;J. Ramanujam

  • Affiliations:

  • Pennsylvania State University, University Park, PA;Pennsylvania State University, University Park, PA;Pennsylvania State University, University Park, PA;Pennsylvania State University, University Park, PA;Pennsylvania State University, University Park, PA

  • Venue:
  • OM '01 Proceedings of the 2001 ACM SIGPLAN workshop on Optimization of middleware and distributed systems
  • Year:
  • 2001

Quantified Score

Hi-index 0.00

Visualization

Abstract

Computer architects have tried to mitigate the consequences of high memory latencies using a variety techniques. An example of these techniques is multi-level caches to counteract the latency that results from having a memory that is slower than the processor. Recent research has demonstrated that compiler optimizations that modify data layouts and restructure computation can be successful in improving memory system performance. However, in many cases, working with a fixed cache configuration prevents the application/compiler from obtaining the maximum performance. In addition, prompted by demands in portability, long battery life, and low-cost packaging, the computer industry has started viewing energy and power as decisive design factors, along with performance and cost. This makes the job of the compiler/user even more difficult as one needs to strike a balance between low power/energy consumption and high performance. Consequently, adapting the code to the underlying cache/memory hierarchy is becoming more and more difficult.In this paper, we take an alternate approach and attempt to adapt the cache architecture to the software needs. We focus on array-dominated applications and measure the potential benefits that could be gained from a morphable (reconfigurable) cache architecture. Our results show that not only different applications work best with different cache configurations, but also that different loop nests in a given application demand different configurations. Our results also indicate that the most suitable cache configuration for a given application or a single nest depends strongly on the objective function being optimized. For example, minimizing cache memory energy requires a different cache configuration for each nest than an objective which tries to minimize the overall memory system energy. Based on our experiments, we conclude that fine-grain (loop nest-level) cache configuration management is an important step for a solution to the challenging architecture/software tradeoffs awaiting system designers in the future.