On 1/14/2015 12:28 PM, Mike Schwab wrote:
(So it looks like any individual task that is CPU bound and limited to one core will take twice as long.)
What you state is always true for temporal multithreading, which timeslices the execution of CP x with CP y on a single core.
But for simultaneous multithreading it would be a worst case boundary condition, but not the most likely result. There is almost 'dead' time in the execution of instructions while the core must wait for storage to be fetched or stored away. SMT takes advantage of that time, with the best case boundary condition being that no perceptible change occurs in the CPU time required to run two tasks on the same core.
Your mileage will vary program by program, but typical results might be that two tasks will execute a single core at 70% of their non-SMT instruction rate, but the same core will execute 40% more total instructions between the two threads.
Greg ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
