Workload Characterization for Branch Predictability

Workload Characterization for Branch Predictability Vikas Texas A&M University Email - vikas.dce2016@gmail.com Paul Gratz Texas A&M University Daniel A. Jim´enez Texas A&M University Abstract—Conditional branch prediction predicts the likely direction of a conditional branch instruction to support ILP ex- traction. Branch prediction is a pattern recognition problem that learns mappings between a context to the branch outcome. An accurate predictor reduces the number of instructions executed on the wrong path resulting in an improvement of performance and energy consumption. In this paper, we present a workload characterization methodology for branch prediction. We propose two new workload-driven branch prediction accuracy identifiers – branch working set size and branch predictability. These parameters are highly correlated with misprediction rates of modern branch prediction schemes (e.g. TAGE and perceptron). We define the branch working set of a trace as a group of most frequently occurring branch contexts, i.e., the 3-part tuple of branch address, and associated global and local history. We analyze the branch working set’s size and predictability on a per-trace basis to study its relationship with a modern branch predictor’s accuracy. We have characterized 2,451 workload traces into seven branch working set size and nine predictability categories after analyzing their branch behavior. We present further insights into the source of prediction accuracy and favored workload categories for modern branch predictors. Index Terms—branch prediction, working set, predictability I. INTRODUCTION Branch prediction is a primary contributor to performance in modern processors. Programs have multiple branch in- structions that can change the direction of the program flow. Branch outcomes are unknown for many cycles. However, the processor must be continually fed to exploit the instruction- level parallelism present in a program. Thus, branch predictors are employed to predict the branch outcome and improve the flow of instructions through the pipeline. Branch prediction leads to instructions being speculatively executed on the predicted path. If the predict is correct, execution continues uninterrupted. Otherwise, the pipeline must be flushed to get rid of wrong-path instructions. Based on pipeline depth and instruction window size, a branch misprediction can lead to large penalties resulting in wasted time and energy. An ideal branch predictor would remember the behavior of all previous branches and the relationships between those branches to the current branch to give very low mispre- diction rates. However, a modern branch predictor with a *Note: This manuscript is an archival version of work conducted as part of the author’s 2020 Master’s thesis at Department of ECEN, Texas A&M University, College Station, TX under the supervision of Professors Paul Gratz and Daniel A. Jim´enez. This submission reflects the same technical content as the manuscript previously submitted to IISWC 2020 and ISPASS 2021. All research in this submission was performed entirely during the author’s time as a graduate student at Texas A&M University. No part of this work was conducted at, funded by, or related to the author’s current employer. limited hardware budget does not always have high accu- racy. Mispredictions can occur due to multiple reasons – too many independent branch contexts to remember within the restricted hardware budget, no recognizable pattern, or the patterns rooted too deeply in history. Misprediction rate and the associated penalty determine the impact on a processor’s performance. Much previous work has precisely quantified this misprediction penalty [1]. However, to the best of our knowl- edge, there have been very few attempts to analyze the mispre- diction rate patterns in workloads and form a branch prediction based workload-characterization methodology. We note that past Championship Branch Prediction (CBP) competitions compare the designs for their performance in application-based workloads groups. This type of comparison serves the purpose of finding out an overall superior branch predictor. However, it gives us no further insight into the relationship between a predictor’s accuracy and branch characteristics of a workload, and whether a design prefers a particular branch distribution. Applications are evolving at a rapid pace so it is imperative that we understand a workload’s inherent branch behavior and predictability to estimate the scope of improvement in the current state of the art in branch predictor designs. Further, a deeper understanding of the sources of branch misprediction could point to new directions in branch prediction. In this work, we show that a branch predictor’s accuracy can be judged by identifying and studying a suitable subset of branch contexts in the workload, called the branch working set (BWSET). We note that a tuple of a branch instruction’s program counter (PC) and prior branch

This content is AI-processed based on open access ArXiv data.