Research on Fast Execution using Helper Threads on Multi-threading Processors

最近，市场上出现了许多多核处理器。在本研究中，我们研究了如何利用多线程技术在这些多核处理器上加速程序。在2005财年，我们对这一研究领域进行了调查，并对算法及其应用进行了研究。在算法方面，我们研究了如何减少由于分支预测错误而导致的CPU管道开放槽数。此外，由于许多多核cpu共用一个二级缓存，我们研究了如何控制二级缓存来加速程序，而如何控制二级缓存是加速程序的关键。通过提出将数据放在缓存中最适合的区域以减少布线延迟的技术，我们已经证实该技术能够使SPECint95/SPECint200程序的IPC平均提高1.17倍。在应用研究方面，我们研究了许多适合多核cpu的应用，包括与搜索相关的应用。在2006财年，本研究的目标是基于2005财年的研究，更多地关注如何加速磁盘访问。我们研究了三个方面：1)从磁盘预加载数据；2)将磁盘缓存扩展到其他机器的内存；3)通过并行磁盘访问来加速shell脚本。首先，我们提出了一种使用辅助线程从磁盘预加载的新方案。使用“gzip”应用程序的实验评估表明，与使用无帮助线程相比，我们可以归档39.2%的加速。其次，我们提出了一种利用另一台PC机的远程内存进行磁盘缓存的新方案。该方案通过使用helper线程实现磁盘缓存的扩展。经证实，在名为DBT-3的基准程序上，加速比高达3.08。第三，我们提出了一种新的并行方案，即在shell脚本程序中使用替代线程。通过使用我们的方案，shell脚本程序的执行速度比正常执行速度提高了1.4到1.8倍。最后的结果现在由USP实验室产品化（http://www.usp-lab.com/）。少

Recently, many multi-core processors are on markets. In this research, we have studied how to accelerate programs on these multi-core processors by using multi-threading technique.In 2005FY, we surveyed this research area, and studied both on the algorithms and on its applications. As for the study on algorithms, we studied how to decrease the number of open slots of CPU pipelines resulted from branch miss-predictions. Moreover, we studied how to control L2 cache to speedup programs because many multi-core CPUs have shared a L2 cache and how to control L2 cache is the key to speedup programs. By proposing the technique that puts data on the best suited area on the cache to decrease wiring delay, we have confirmed that the technique is able to speedup SPECint95/SPECint200 programs 1.17 times of IPC in average. As for the study on applications, we investigated many applications including search related ones that will be suited for multi-core CPUs.In 2006FY, the targets of this research w … More ere fixed on how to speedup disk accesses based on 2005FY studies. We researched on three areas : 1) pre-loading of data from disks, 2) extending disk cache onto other machine's memory and 3) shell script acceleration by parallelizing disk access. First, we have proposed a new pre-loading scheme from disks by using helper threads. Experimental evaluation using "gzip" application shows that we can archive 39.2% speedup in comparison with using no-helper threads. Second, we have proposed a new disk caching scheme by using remote memory of another PC. The scheme realize extending the size of disk cache by using helper-threads. It has been confirmed that the speedup ratio becomes up to 3.08 on the benchmark program called DBT-3. Third, we have proposed a new parallelizing scheme by using alternative threads for shell script programs. By using our scheme, shell script programs archive 1.4 to 1.8 times speedup in comparison with normal execution. The last result is now under productization by USP Lab (http://www.usp-lab.com/). Less