So basically, on the Architecture Moed Bet exam, they gave us this branch-prediction problem. And what they WANTED was for us to consider the issue of how well variously local and global predictor types will predict when predicting randomly on programs that compute whether to take their jumps randomly. What they WANTED was very simplistic probability calculations. Unfortunately, the problem they WROTE would be CORRECTLY solved by considering the branch predictor as a Markov Decision Process with transitions defined by the probability P of actually branching, writing out the transition matrix for that MDP, and finding its steady-state distribution (the matrixs eigenvector, normalized to sum to 1.0). In all cases, there were 4 entries in the Branch Target Buffer, and for the local-local case there were 4 state-machine entries in each BTB entry which also had its own Branch History Register. So basically, the correct way to do it was to calculate out these 20 or so massive MDPs analytically, and then plug in three different values of P, and THEN be able to tell how well the branch predictor does on each of three randomized conditional branches. But what they WANTED was to pretend the predictor gives a random answer, to thwart the students ability to just simulate the predictor in his mind or on-paper, because God forbid students take the Calculemus! approach on exams. Goddamnit Technion.
Posted on: Tue, 07 Oct 2014 12:23:44 +0000
Trending Topics
Recently Viewed Topics
© 2015