Performance and queueing theory
If we observe the performance benchmark numbers across a number of runs, even though the hardware, loads and OS remain the same, the numbers are rarely exactly the same. The difference between each run may be as much as -8 percent to 8 percent for no apparent reason. This may seem surprising, but the deep-rooted reason is that the performances of computer systems are stochastic by nature. There are many small factors in a computer system that make performance unpredictable at any given point of time. At best, the performance variations can be explained by a series of probabilities over random variables.
The basic premise is that each subsystem is more or less like a queue where requests await their turn to be served. The CPU has an instruction queue with unpredictable fetch/decode/branch-predict timings, the memory access again depends on cache hit ratio and whether it needs to be dispatched via the interconnect, and the I/O subsystem works using interrupts...
