S. Sechrest, C. C. Lee, T. Mudge. "The Role of Adaptivity in Two-level Adaptive Branch Prediction". 28 th International Symposium on Microarchitecture. 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....effects. If such histories are to be used as a measure of confidence, it is necessary to know which ones are (normally) followed by a successful prediction and which ones are not. Heuristics and algorithms to do this exist in the branch prediction literature. For instance, Sechrest et al. [28] describe a scheme that tries to identify repeating patterns of branch outcomes. If no repeating pattern can be detected, a simple population count is used. They call this scheme algo. As an alternative, Sechrest et al. suggest running a set of programs and recording the behavior. They call this ....

....the PHT entries can be modified onthe fly. The third (lowercase) letter is identical to the first letter, except it describes the PHT instead of the BHR. Hence, g means one global PHT, s means one PHT per set, and p means one PHT per address. Sechrest, Lee, and Mudge extend this nomenclature [28]. They distinguish between two ways of programming the fixed entries of the non adaptive PHT schemes. Profile based is denoted by appending comp to the three letters and algorithm based is denoted by appending algo. We adopt the SSg(comp) design for use as a confidence estimator in our load value ....

S. Sechrest, C. C. Lee, T. Mudge. "The Role of Adaptivity in Two-level Adaptive Branch Prediction". 28 th International Symposium on Microarchitecture. 1995.

....the oldest bit is lost. If such histories are to be used as a measure of confidence, it is essential to know which ones are (frequently) followed by a correct prediction and which ones are not. The branch prediction literature describes algorithms to accomplish this. For instance, Sechrest et al. [SLM95] suggest profiling a set of programs to record the behavior. Table 5.2 shows the average SPECint95 last value predictability following each of the sixteen possible four bit prediction outcome history. For example, the second row of the table states that a failure, failure, failure, success ....

S. Sechrest, C. C. Lee, T. Mudge. "The Role of Adaptivity in Two-level Adaptive Branch Prediction". 28 th International Symposium on Microarchitecture. 1995.

....Burtscher Zorn 3 PFDC 98 of related work exists. Yeh and Patt [YePa92, YePa93] describe sets of two level branch predictors and invent a taxonomy to distinguish between them. We adopt one of their designs for use as a confidence estimator in our load value predictor. Sechrest, Lee, and Mudge [SLM95] refine some of Yeh and Patt s two level predictors by studying and describing how to program the static predictor components. They distinguish between profile based and algorithmic approaches. We discuss both schemes in Section 5.1. 4. Methodology All our measurements are performed on the DEC ....

....finer granularity, and they yield much higher accuracies. If such histories are to be used as a measure of confidence, it is necessary to know which ones are (normally) followed by a correct prediction and which ones are not. Heuristics and algorithms to do this exist. For example, Sechrest et al. [SLM95] describe a scheme that first looks for repeating patterns and then falls back to population counting if none can be detected. They named this scheme (algo) As an alternative, they suggest running a set of programs and recording their behavior. This profile based approach is called (comp) We ....

S. Sechrest, C. C. Lee, T. Mudge. "The Role of Adaptivity in Two-level Adaptive Branch Prediction". 28 th International Symposium on Microarchitecture. 1995.

....to the chapters in this dissertation will be discussed within those chapters. Almost all mechanisms include a table of two bit saturating counters the PHT and most work has focused on improving the behavior of this table by considering whether the PHT should in fact use dynamic counters [83], and by showing the importance of aliasing in the PHT as a contributor to the misprediction rate [57, 66, 82, 94] Kim and Tyson [50] examine the impact of aliasing in the table of per branch histories the BHT and find it is a further important contributor of CHAPTER 1. INTRODUCTION 16 ....

S. Sechrest, C.-C. Lee, and T. Mudge. The role of adaptivity in two-level adaptive branch prediction. In Proceedings of the 28th Annual International Symposium on Microarchitecture, pages 264--69, Dec. 1995.

....being correlated in the source code of the SPEC89 benchmarks. Chang et al. 1] classified branches based on their taken rates. They proposed a predictor using a static predictor for the strongly biased branches, and a dynamic hybrid predictor for the weakly biased branches. Sechrest et al. [5] studied the role of adaptivity in twolevel branch predictors and determined that, for per address predictors with short histories, having statically determined values in the PHT performed on par with the adaptive scheme using 2 bit counters. Talcott et al. 9] and Young et al. 12] studied and ....

S. Sechrest, C.-C. Lee, and T. Mudge, "The role of adaptivity in two-level adaptive branch prediction," in Proceedings of the 28th Annual ACM/IEEE International Symposium on Microarchitecture, 1995.

....execution states by using both the branch address and the branch history. Lee and Smith [6] proposed a scheme where the value of each Pattern History Table entry is determined statically, using profile information; this scheme is referred to as the PSg scheme by Yeh and Patt [16] Sechrest et al. [12] introduced another method, PSg (algo) of statically determining the values in the PHT and showed that a statically determining table could perform nearly as well as an adaptive PHT for short branch histories. Since the contents of PHT are determined statically, the PSg scheme trades the benefits ....

....of two bit counters. 2 12 GAs(m,n) the global variation of the Two Level Adaptive Branch Predictor [15] consisting of a single m bit global branch history and n pattern history tables. m 2 m 1 n PSg(m) a modified version of the per address variation of the Two Level Adaptive Branch Predictor [16, 12] consisting of 2K m bit branch history registers and a single pattern history table (each PHT entry uses one statically determined hint bit instead of a 2bC) The version of PSg used in this study is the PSg(algo) 12] 2 11 m 2 m gshare(m) a modified version of the global variation of the ....

[Article contains additional citation context not shown here]

S. Sechrest, C.-C. Lee, and Trevor Mudge, "The Role of Adaptivity in Two-Level Adaptive Branch Prediction," 28th ACM/IEEE International Symposium on Microarchitecture, Nov. 1995.

....biased branches from weakly biased branches based on profiling information. Using static prediction on the strongly biased branches and dynamic prediction on the remaining branches resulted in lower contention in the pattern history table and thus more accurate predictions. Sechrest et al. [3] further classified the branches and showed that a significant number of the biased branches are always taken and always nottaken. In this paper, we propose a new branch predictor that dynamically classifies branches based on their history patterns. In order to reduce contention, this predictor ....

....higher performance, the disadvantage of having less accurate predictions for these branches must not outweigh the benefits of reducing the PHT interference. For branches that have a recurring pattern, we expect the next outcome of the branch to be an extension of this pattern. Since the PSg(algo) [3] works on this premise, we expect it to perform well for branches with repeating patterns and apply it to these branches. Thus, we evaluate its performance. Figure 3 and 4 show the performance of PSg(algo) with a 1024 and a 8192 entry BTB respectively. Branches with short repeating patterns were ....

S. Sechrest, C.-C. Lee, and T. Mudge. The role of adaptivity in two-level adaptive branch prediction. In Proceedings of the 28th Annual ACM/IEEE International Symposium on Microarchitecture, 1995.

....the regularity seen so far; it appears that it is always beneficial to filter out such events. In other words, the patterns currently stored should only updated by new patterns, and a single event is not a good enough indication of the start of a new pattern (others have found similar effects [SLM95], J 96] It may be interesting to look at alternative rules such as predictors that only change an association after three consecutive misses, or only if the target of the second miss was identical to the first one. 6. Related work Lee and Smith [LS84] describe several forms of BTBs, including ....

Stuart Sechrest, Chich-Chieh Lee, and Trevor Mudge. The role of adaptivity in two-level adaptive branch prediction. Proceedings of MICRO-29, November 1995.

....benchmarks, shown in Figure 13 and 14. As the optimal points move downward (budgets increase) the number of history bits increases (left numbers in the parentheses) We also notice that the number of history bits increase faster for SPEC, because branches in SPEC are relatively harder to predict [Sechrest95]. Also note that as budgets increase, the number of entries in the branch history table also increases with the number of history bits. This trend can be explained by the increasing importance of history information. As described in the previous paragraph, history information becomes important ....

Sechrest, S., Lee, C-C. and Mudge, T. The role of adaptivity in two-level adaptive branch prediction. Proceedings of the 28th Annual International Symposium on Microarchitecture, 264269, November 1995.

....and used speculatively pending actual computation of the value. Previous work on categorizing and characterizing predictors has been based on the high level constructs from earlier research on branch prediction mechanisms, such as global history components or tables of saturating counters [12,13,18,19]. This seems logical, particularly for an automated search, because a predictor composed from such constructs is quite likely to resemble a known good predictor, whereas a predictor built arbitrarily from lower level primitives is less likely to be useful. Using higher level constructs, however, ....

....predictor, the choosing predictor [10] used in the DEC Alpha 21264. This predictor consists of a local history predictor and a global history predictor that are selected between by a global history based chooser: Lhist[ PC;T) Threebit[1K] Hist[10,1K] PC;T) T) Ghist[ T) Twobit[4K] Hist[12,1](0;T) T) A21264[ PC; T) if (MSB(P[2,4K] Hist[12,1] 0;T) P (Lhist= T) Ghist = T) then Lhist[ PC;T) else Ghist[ T) Where Threebit is the obvious extension of Twobit to three bit counters. Through the composition of predictors and various logic expressions a large variety of ....

[Article contains additional citation context not shown here]

S. Sechrest, C-C Lee and T. Mudge, "The Role of Adaptivity in Two-level Adaptive Branch Prediction", 28th ACM/IEEE International Symposium on Microarchitecture, Nov 1995.

....Weighted average 9 87.47 94.84 98.55 95.17 88.62 92.930 94.056 15 89.69 96.43 98.65 96.84 96.27 95.577 97.118 20 90.23 96.74 98.69 97.28 96.83 95.952 97.418 January 31, 1996 Limits to Branch Prediction 14 no variance. For calibrating prediction algorithms, a set of inputs would be more desirable [Sechrest95]. 6. Conclusions and further work In this paper, we assert that there is a fundamental limit to branch predictability, due to the randomness of the input data. Knowing or estimating this limit before attempting to design branch predictors is valuable. If there is a large gap between the ....

Sechrest, S., Lee, C. and Mudge, T. The role of adaptivity in two-level adaptive branch prediction. Proceedings of the 26th Annual International Symposium on Microarchitecture, Ann Arbor, December 1995.

....13 history bits achieves a predication accuracy of approximately 92 . Lee and Smith [17] proposed a scheme where the value of each Pattern History Table entry is determined statically, using profile information; this scheme is referred to as the PSg scheme by Yeh and Patt [40] Sechrest et al. [31] introduced another method, PSg (algo) of statically determining the values in the PHT. The PSg (algo) works on the premise that for branches with a recurring pattern, the next outcome of the branch is likely an extension of this pattern. For example, if the branch history of a branch is ....

....to achieve higher performance, the disadvantage of having less accurate predictions for these branches must not outweigh the benefits of reducing the PHT interference. For branches that have a recurring pattern, the next outcome of the branch is an extension of this pattern. Since the PSg(algo) [31] works on this premise, we expect it to perform well for branches with repeating patterns and apply it to these branches. Figure 6.4 compares the performance of PSg(algo) to that of gshare. Branches with 1 2 3 4 5 6 7 8 Pattern Length 0.0 0.1 0.2 Misprediction Rate PSg gshare (2KB) gshare ....

S. Sechrest, C.-C. Lee, and T. Mudge, "The role of adaptivity in two-level adaptive branch prediction," in Proceedings of the 28th Annual ACM/IEEE International Symposium on Microarchitecture, 1995.

No context found.

Stuart Sechrest, Chich-Chieh Lee, and Trevor Mudge. The role of adaptivity in two-level adaptive branch prediction. Proceedings of MICRO-29, November 1995.

No context found.

S. Sechrest, C. C. Lee, T. Mudge. "The Role of Adaptivity in Two-level Adaptive Branch Prediction ". 28 th International Symposium on Microarchitecture. 1995.

No context found.

Stuart Sechrest, Chich-Chieh Lee, and Trevor Mudge. The role of adaptivity in two-level adaptive branch prediction. Proceedings of MICRO-29, November 1995.

No context found.

S. Sechrest, C.-C. Lee, and T. Mudge, "The Role of Adaptivity in Two-Level Adaptive Branch Prediction, " in Proceedings of the 28th International Symposium on Microarchitecture (MICRO-28), pp. 264-- 269, IEEE and ACM, November/December 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC