Research On Quality-of-Service Guarantee For Network-on-Chip Based On Application-Aware Method

With the increase of the size of user number and machine server, Internet companies tend to migrate their service to the cloud. On the data center, one server may run a number of different applications. These applications due to different business functions, the operating characteristics of the program is also very different, and the resource demand and usage are not the same either. In particular, the use of shared resources makes sense. The application with low resource requirements may by interfered by applications with high resource usage. If we take the same treatment strategy to both two, the performance will be seriously affected. Therefore, how to ensure quality of service for different applications in a mixed load, becoming the hot topic of research in multi-core processors.The on-chip network in multi-core processors is an important shared resource, which have the responsibility of communication among core, Cache and DRAM. The quality of communication is depending on the usage of link and buffer. Currently the link usage that based on application-aware is mainly based on the miss rate of Last Level Cache to classify applications, which ignores the real relevance between performance and miss rate. The performance is not necessarily decreased when the miss rate increases, so these schemes are inaccurate. While the buffer usage scheme that based on application-aware is to design a heterogeneous network with two kinds of routers, which not work well at high load and unbalanced load.To solve the above problems, this paper proposes a new application-aware bufferless routing(APA) based on the direct network features, and a congestion control mechanism for heterogeneous network design(COP). APA algorithm statistics the average duration of an application's network stage and computing stage and divide the application into delay-sensitive(LS) and delay-insensitive(LIS). When LS has the same priority with LIS, delay has greater impact on the performance of the LS. Therefore, the priority policy of APA also adds the application type as the evaluation parameter to ensure the performance of the LS. As the same time, due to the priority policy retains the parameter of original maximum life span(Age), the LIS's performance will not be affected. With the growing popularity of large-load applications, this paper proposed congestion control in heterogeneous No C, which determines whether the current No C needs resource or feels free to help the other No C by detecting the congestion degree in different No C. This approach solves the serious congestion problem in isolated network under large workload and unbalancing workload. Combined with application-aware policy, COP ensures the applications performance in mixed loads.This paper used Mac Sim as the simulation platform, based on 64-core chip. APA enhance the system performance by average of 4.9%(up to 10.4%), delay reduction by average of 2.4% and deflection reduction by average of 4.2%, compared to the nonapplication-aware scheme. COP scheme have performance improvements of 17% by average(up to 24%), compared to the baseline that without congestion control, and have a power saving more than 36% than 64-core pure No C. Therefore, the APA scheme and COP scheme guarantee a good quality of service improvement.