{"title":"减少嵌入式多处理器并行处理开销的体系结构支持","authors":"Jian Wang, Joar Sohl, Dake Liu","doi":"10.1109/EUC.2010.17","DOIUrl":null,"url":null,"abstract":"The host-multi-SIMD chip multiprocessor (CMP) architecture has been proved to be an efficient architecture for high performance signal processing which explores both task level parallelism by multi-core processing and data level parallelism by SIMD processors. Different from the cache-based memory subsystem in most general purpose processors, this architecture uses on-chip scratchpad memory (SPM) as processor local data buffer and allows software to explicitly control the data movements in the memory hierarchy. This SPM-based solution is more efficient for predictable signal processing in embedded systems where data access patterns are known at design time. The predictable performance is especially important for real time signal processing. According to Amdahl¡¯s law, the nonparallelizable part of an algorithm has critical impact on the overall performance. Implementing an algorithm in a parallel platform usually produces control and communication overhead which is not parallelizable. This paper presents the architectural support in an embedded multiprocessor platform to maximally reduce the parallel processing overhead. The effectiveness of these architecture designs in boosting parallel performance is evaluated by an implementation example of 64x64 complex matrix multiplication. The result shows that the parallel processing overhead is reduced from 369% to 28%.","PeriodicalId":265175,"journal":{"name":"2010 IEEE/IFIP International Conference on Embedded and Ubiquitous Computing","volume":null,"pages":null},"PeriodicalIF":0.0000,"publicationDate":"2010-12-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"5","resultStr":"{\"title\":\"Architectural Support for Reducing Parallel Processing Overhead in an Embedded Multiprocessor\",\"authors\":\"Jian Wang, Joar Sohl, Dake Liu\",\"doi\":\"10.1109/EUC.2010.17\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"The host-multi-SIMD chip multiprocessor (CMP) architecture has been proved to be an efficient architecture for high performance signal processing which explores both task level parallelism by multi-core processing and data level parallelism by SIMD processors. Different from the cache-based memory subsystem in most general purpose processors, this architecture uses on-chip scratchpad memory (SPM) as processor local data buffer and allows software to explicitly control the data movements in the memory hierarchy. This SPM-based solution is more efficient for predictable signal processing in embedded systems where data access patterns are known at design time. The predictable performance is especially important for real time signal processing. According to Amdahl¡¯s law, the nonparallelizable part of an algorithm has critical impact on the overall performance. Implementing an algorithm in a parallel platform usually produces control and communication overhead which is not parallelizable. This paper presents the architectural support in an embedded multiprocessor platform to maximally reduce the parallel processing overhead. The effectiveness of these architecture designs in boosting parallel performance is evaluated by an implementation example of 64x64 complex matrix multiplication. The result shows that the parallel processing overhead is reduced from 369% to 28%.\",\"PeriodicalId\":265175,\"journal\":{\"name\":\"2010 IEEE/IFIP International Conference on Embedded and Ubiquitous Computing\",\"volume\":null,\"pages\":null},\"PeriodicalIF\":0.0000,\"publicationDate\":\"2010-12-11\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"5\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"2010 IEEE/IFIP International Conference on Embedded and Ubiquitous Computing\",\"FirstCategoryId\":\"1085\",\"ListUrlMain\":\"https://doi.org/10.1109/EUC.2010.17\",\"RegionNum\":0,\"RegionCategory\":null,\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"\",\"JCRName\":\"\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"2010 IEEE/IFIP International Conference on Embedded and Ubiquitous Computing","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1109/EUC.2010.17","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"","JCRName":"","Score":null,"Total":0}
Architectural Support for Reducing Parallel Processing Overhead in an Embedded Multiprocessor
The host-multi-SIMD chip multiprocessor (CMP) architecture has been proved to be an efficient architecture for high performance signal processing which explores both task level parallelism by multi-core processing and data level parallelism by SIMD processors. Different from the cache-based memory subsystem in most general purpose processors, this architecture uses on-chip scratchpad memory (SPM) as processor local data buffer and allows software to explicitly control the data movements in the memory hierarchy. This SPM-based solution is more efficient for predictable signal processing in embedded systems where data access patterns are known at design time. The predictable performance is especially important for real time signal processing. According to Amdahl¡¯s law, the nonparallelizable part of an algorithm has critical impact on the overall performance. Implementing an algorithm in a parallel platform usually produces control and communication overhead which is not parallelizable. This paper presents the architectural support in an embedded multiprocessor platform to maximally reduce the parallel processing overhead. The effectiveness of these architecture designs in boosting parallel performance is evaluated by an implementation example of 64x64 complex matrix multiplication. The result shows that the parallel processing overhead is reduced from 369% to 28%.