Analysis of False Cache Line Sharing Effects on Multicore CPUs

False sharing (FS) is a well-known problem occurring in multiprocessor systems. It results in performance degradation on multi-threaded programs running on multiprocessor environments. With the evolution of processor architecture over time, the multicore processor is a recent direction used by hardware designers to increase performance while avoiding heat and power walls. To fully exploit the processing power from these multicore hardware architectures, the software programmer needs to build applications using parallel programming concepts, which are based upon multi-threaded programming principles. Since the architecture of a multicore processor is very similar to a multiprocessor system, the presence of the false sharing problem is speculated. Its effects should be measurable in terms of efficiency degradation in a concurrent environment on multicore systems. This project discusses the causes of the false sharing problem in dual-core CPUs, and demonstrates how it lessens the system performance by measuring efficiency of a test program in sequential compared to parallel versions. Thus, demonstration programs are developed to read a CPU cache line size, and collect the execution results of the test program with and without false sharing on the specific system hardware. Certain techniques are implemented to eliminate false sharing. These techniques are described, and their effectiveness in mitigating the speed-up and efficiency lost from false sharing is analyzed.False sharing (FS) is a well-known problem occurring in multiprocessor systems. It results in performance degradation on multi-threaded programs running on multiprocessor environments. With the evolution of processor architecture over time, the multicore processor is a recent direction used by hardware designers to increase performance while avoiding heat and power walls. To fully exploit the processing power from these multicore hardware architectures, the software programmer needs to build applications using parallel programming concepts, which are based upon multi-threaded programming principles. Since the architecture of a multicore processor is very similar to a multiprocessor system, the presence of the false sharing problem is speculated. Its effects should be measurable in terms of efficiency degradation in a concurrent environment on multicore systems. This project discusses the causes of the false sharing problem in dual-core CPUs, and demonstrates how it lessens the system performance by measuring efficiency of a test program in sequential compared to parallel versions. Thus, demonstration programs are developed to read a CPU cache line size, and collect the execution results of the test program with and without false sharing on the specific system hardware. Certain techniques are implemented to eliminate false sharing. These techniques are described, and their effectiveness in mitigating the speed-up and efficiency lost from false sharing is analyzed