Modeling and Performance Estimation for Multiprocessor System On Chip Architectures

Nowadays mobile operators face several challenges because of the large growth of internet traffic usage, and the increase of the number of users. There is a need for innovations in almost every area such as e-health, private businesses, automation, etc. 5G supports a variety of new services with different requirements for throughput, latency, and reliability. Multicore computing platforms are used to meet the various implementations while allowing scalability and flexibility in the implementation of the base stations. The challenge in this regard is the efficient distribution and processing of signal processing tasks on parallel processors. Multiple processing elements give an opportunity to exploit application parallelism by splitting them into many parallel tasks and making the parallel execution possible. But unfortunately, these methods also have disadvantages. Moreover, with increasing the application complexity, the management and synchronization overhead increases disproportionately, which limits the increase in performance and system efficiency. The performance of the application highly depends on the methods used to execute it. The overheads resulting from synchronization, management, and waiting time in a queue sometimes are not solvable only by choosing the right scheduling algorithm. Therefore, another solution is needed. To cope with this problem, the application granularity reduction using task clustering was proposed recently and demonstrated impressive performance improvement. Clustering application before applying scheduling algorithm reduces the granularity of the application and solves the problem concerning synchronization and management overheads. There are different clustering algorithms to be chosen, but they are non deterministic and have high interconnection between tasks from various clusters. Our motivation for this work was to find a clustering algorithm that will suit our problem better and will result in better performance. We present a modification of the CASS-II clustering algorithm proves to be a satisfactory solution for our applications. In order to analyze and simulate different graphs, we also developed a simulation tool, which enables us to run different algorithms for scheduling and clustering very efficiently. It accepts any kind of graphs, and the results can be later analyzed via easily formed visualization functions. We also covered another important topic that is energy consumption. Power control is one of the most important topics in many communication and computation environments. Heat generation, expensive packing, and cooling can be a result of high power generation. Many researchers are engaged in this problem and suggest various solutions. Energy-delay tradeoff is examined as a method for energy saving. In this work, we investigate dynamic voltage and frequency scaling and suggest a new modified method called Proportional Task Scaling. In order to change energy consumption, the frequency and voltage should be changed together. There are different algorithms that suggest reducing these parameters differently. The greedy static power management algorithm suggests to reduce frequency and increase the execution time of the first tasks on processing elements. Most of the algorithms use the slack value(difference of deadline and scheduling length) to make this change happen. Our motivation is to reduce even more energy consumption by still meeting the deadline of the application. So, our suggestion is to change parameters not only for the first but the whole application by meeting scheduling requirements and deadlines.