The choice between using locks and transactional memory (TM) semantics is not always obvious, particularly given the hardware complexity of implementing the TM semantics. This exercise illustrates...

The choice between using locks and transactional memory (TM) semantics is not always obvious, particularly given the hardware complexity of implementing the TM semantics. This exercise illustrates these trade-offs. Consider that two threads T1 and T2 are executing two pieces of code that may need mutual exclusion, since the two threads conditionally modify a shared variable. Each thread executes ten instructions in their respective critical section. If the developer uses locks, the code can run on machine A. If the developer uses transactions, the code can run on machine B. Both A and B are multi-threaded machines that allow T1 and T2 to be executed concurrently. Both machine A and machine B can execute the critical section codes with a CPI of 1. The only difference between the two machines is that machine B’s cycle time is 20% longer than that of machine A as it has to support TM semantics. If a transaction fails, there is an additional penalty of 20 cycles to clear the transactional state.

(a) If the probability that the two threads concurrently read/modify the shared variable 10% of the time (i.e., 90% of the time there is no sharing), which machine is faster?

(b) As the size of the critical section grows, the probability that there is a conflict increases. Hence, if the number of instructions in the critical section increases from 10 to 1000, the probability of contention increases from 10% to 20%. In such a scenario, which machine is faster?