| 摘要 |
A method, system, and computer program product are provided for profiling virtual memory accesses by one or more threads. A virtual memory access thread profiling tool includes a histogram generator and a thread placement file generator. The histogram generator generates a histogram that indicates the relative frequency at which virtual memory addresses are accessed by each program thread. To generate the histogram, the histogram generator runs and interrupts each program thread to collect samples. When an interrupt is issued, a program counter is returned. A valid load or store instruction is determined for a thread in assembly code identified by the returned program counter. In one example, to determine a valid load or store instruction, the histogram generator walks forward or backward through the assembly code identified by the returned program counter until a valid load or store instruction is reached. A virtual memory address corresponding to a valid load or store instruction is then read. A histogram is then incremented based on the virtual memory address to track the relative frequency at which virtual memory addresses are accessed by each thread. Further options are provided to allow a user to control the virtual memory access profiling tool. These options include selecting and setting the type of collection method, selecting and setting program counter (PC) and memory address filtering, setting the sampling granularity, providing start/stop control, selecting real-time graphical histogram output, and naming output files. The virtual memory access thread profiling tool can be implemented as a profiling command executed by an operating system for a multi-processor system, such as, a NUMA machine. |
