Application-Specific Hardware-Driven Prefetching to Improve Data Cache Performance

• Published in: Advances in Computer Systems Architecture pp. 761 - 774
• Main Authors: Modarressi, Mehdi, Goudarzi, Maziar, Hessabi, Shaahin
• Format: Book Chapter Conference Proceeding
• Language: English
• Published: Berlin, Heidelberg Springer Berlin Heidelberg 2005; Springer
• Series: Lecture Notes in Computer Science
• Subjects: Applied sciences; Cache Block; Cache Size; Computers, microcomputers; Electronics; Exact sciences and technology; Hardware; Integrated circuits; Integrated circuits by function (including memories and processors); Prefetch Instruction; Prefetching Scheme; Prefetching Technique; Semiconductor electronics. Microelectronics. Optoelectronics. Solid state devices; Access time; Prefetching; Cache memory; Time average; Customization; Computer architecture; Object oriented; Program performance; Reconfigurable architectures; Custom circuit
• ISBN: 3540296433; 9783540296430
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/11572961_62

Data cache hit ratio has a major impact on execution performance of programs by effectively reducing average data access time. Prefetching mechanisms improve this ratio by fetching data items that shall soon be required by the running program. Software-driven prefetching enables application-specific policies and potentially provides better results in return for some instruction overhead, whereas hardware-driven prefetching gives little overhead, however general-purpose processors cannot adapt to the specific needs of the running application. In the application-specific processors that we develop customized to an object-oriented application, we implement application-specific hardware prefetching to benefit from both worlds. This prefetching policy prefetches all data items that shall be unconditionally accessed by a class method when the class method is called. We mathematically analyze this mechanism and present its simulation results using some object-oriented benchmarks. Simulation results in absence and presence of the proposed prefetching mechanism confirm the theoretical results and show that on average, the miss ratio is reduced by 73%.