4-iş parçacıklı işlemcilerde servis kalitesi ve verimlilik optimizasyonu

Simultaneous multithreading (SMT) is one of the key performance features in modern processors. However, in data center applications where immense computing power is required, there is a certain aspect that is much more important than overall performance: Quality of service. SMT, though, meanwhile getting its strength from sharing resources between multiple threads in a core, actually hurts the throughput of singular workloads, and a significant amount of these workloads require an exorbitant degree of quality of service. Thus, SMT, which is one of the breakthroughs lying under the modern processor microarchitecture, actually hurts data center applications due to its sharing nature. This handicap even leads the data centers to the point of disabling SMT on their processors to preserve their required quality of service. This topic has been occasionally studied in the literature and several propositions were made to overcome this barrier and obtain higher performance and increase performance per Watt and performance per total cost of ownership dollars. One of the leading studies is QoSMT which aims to enable a quality of service satisfying SMT approach in data centers. It achieves so by implementing a statistic-based dynamic resource partitioning scheme. However, the original QoSMT paper implements a design for two-threaded SMT cores meanwhile the modern data center processors use four or eight threads. When there are only two threads in a core, there is only a single transaction pathway between threads. High priority thread (HPT) takes or gives resources from/to low priority thread. When the thread count increases though, a necessity for resource scheduling arises as well and the statistical accuracy of partition decisions naturally degrades. In this study, we take the baseline QoSMT study, implement it, extend it to four-threaded SMT cores, suggest and implement some overall design improvements and analyze its capabilities for three different scheduling algorithms. We observe that our proposed best designs provided up to 66.63% quality of service improvements compared to the baseline SMT with only a 7.09% loss of total performance in terms of IPC.