Etian Federovsky, Meir Feder, and Shlomo Weiss. Branch prediction based on universal data compression algorithms. In Proceedings of the 25th Annual International Symposium on Computer Architecture, pages 62--72, June 1998.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....used to make the final decision as follows: ffl if ratio 0:5 prediction = 1 (branch is taken) ffl if ratio 0:5 Gamma prediction = 0 (branch is non taken) ffl else prediction = 1 with probability Phi(P ) where = 1 p T and T is the branch history length. The authors in [140] also discuss variations of the CWT algorithm, where they add bits from the branch address in the past branch history, to further increase the levels of the context tree. The original CWT construction algorithm assumes that a complete context tree is built, where the total number of nodes is ....

....context where sufficient prediction information exists. The conditional probability given by b 0:5 a b 1 makes a prediction using the same algorithm as in CWT. An adaptive PPM algorithm can also be derived if we consider generating and deallocating nodes, as in the adaptive CWT algorithm. In [140], idealized versions of these algorithms are applied to conditional branch prediction with some success. However, a realistic hardware implementation has not yet been described. Another version of the PPM algorithm is described in [141] It has been successfully applied to areas such as data ....

E. Federovksy, M. Feder, and S. Weiss. Branch Prediction based on Universal Data Compression Algorithms. In Proceedings of the International Symposium on Architecture, pages 62--72, June 1998. 189

....However, PPM prediction has a predetermined table size for all its components (exponentially decreasing in size) reserves a component for each history bit length, and does not employ filtering (similar to the ideal staged predictor discussed in Section 9.4.3. 2) Federovsky, Feder and Weis [57] also adapt a compression technique, Context Tree Weighting (CWT) to conditional branch prediction. They demonstrate that the technique delivers excellent prediction accuracy under the assumption of unconstrained hardware resources. They do not assess prediction performance under limited hardware ....

E.Federovsky, M.Feder, S.Weiss. Branch Prediction based on Universal Data Compression Algorithms. ISCA `98 Conference Proceedings, pp. 62-72, Barcelona, July 1998

.... We discuss two major compression techniques: the first is the Prediction by Partial Matching (PPM) 26, 27] algorithm that has been proposed for predicting conditional branches in [28] and the second (and most recent) one is the Context Weight Tree (CWT) method [29] which has been presented in [30] as a another potential conditional branch predictor. The primary goal of data compression is to represent an original data sequence with another that has fewer data elements. Modern data compression techniques try to use fewer bits to represent frequent symbols, thus reducing the overall size of ....

....used to make the final decision as follows: ffl if ratio 0:5 prediction = 1 (branch is taken) ffl if ratio 0:5 Gamma prediction = 0 (branch is non taken) ffl else prediction = 1 with probability Phi(P ) where = 1 p T and T is the branch history length. The authors in [30] also discuss variations of the CWT algorithm where they add bits from the branch address in the past branch history, to further increase the levels of the context tree. In addition, the original CWT construction algorithm assumes that a complete context tree is built, where the total number of ....

[Article contains additional citation context not shown here]

E. Federovksy, M. Feder, and S. Weiss. Branch Prediction based on Universal Data Compression Algorithms. In Proceedings of the International Symposium on Architecture, pages 62--72, June 1998.

....Information theory is a related field to computer science and statistics, so it is not surprising that ideas from information theory are relevant to problems in profiling. Some information theory has been used recently to analyze dynamic branch prediction schemes (Chen, Coffey, Mudge, 1996, and Federovsky, Feder, Weiss, 1998). To our knowledge, this is the first work that applies information theoretic concepts to profiling. Profiles are statistics about the execution of a program; they are commonly execution frequencies but they can also include data about system performance at a variety of levels. Our information ....

Federovsky, E., Feder, M., & Weiss, S. (1998). Branch Prediction Based on Universal Data Compression Algorithms. In Proc. Twenty-Fifth International Symposium on Computer Architecture.

....a large feature space. We argued in Section 3 that branch prediction is one such a domain, it is reasonable to expect other speculative prediction domains to have this property. It may appear that decision trees are similar to context trees, which have been previously used for branch prediction [11] and are general enough to be applied to other hardware prediction problems; however, a decision tree can capture a far greater variety of predictive rules than a context tree of the same depth, and context trees exhibit the same exponential growth in the number of features considered that ....

....further size increase. The effect is striking, and is responsible for the advantage PAp has over DDT at large size, but we have not found an explanation at this point. 7 Related Work Here we consider only the work most closely related to our proposed approach but not covered above. The work in [11] introduced the use of context trees for branch prediction however, this predictor d d 1 is not tied to the branch prediction domain and could be applied to other domains with no major changes. The rules that can be expressed by a context tree are a strict subset of the rules that can ....

Etian Federovsky, Meir Feder, and Shlomo Weiss. Branch prediction based on universal data compression algorithms. In Proceedings of the 25th Annual International Symposium on Computer Architecture, pages 62--72, June 1998.

....compression, and prediction. Information theory is a related field to computer science and statistics, so it is not surprising that ideas from information theory are relevant to problems in profiling. Some information theory has been used recently to analyze dynamic branch prediction schemes [5,9]. To our knowledge, this is the first work that applies information theoretic concepts to profiling. Profiles are statistics about the execution of a program; they are commonly execution frequencies but they can also include data about system performance at a variety of levels. Our ....

E. Federovsky, M. Feder, and S. Weiss. "Branch Prediction Based on Universal Data Compression Algorithms," Proc. Twenty-Fifth International Symposium on Computer Architecture. New York: ACM, June-July 1998.

.... processors, such static fall back schemes are used [7] 5 CTW, Adaptive CTW and Adaptive PPM applied to Branch Prediction: Federovsky et al. 1998 Inspired by the observation of the similarity between the prediction of branch outcomes and the predictors used in data compression, Federovsky et al. [6] extended the study begun by Chen et al. 1] In addition to examining the performance of adaptive versions of the PPM algorithm, they also consider static and adaptive versions of the context tree weighting (CTW) algorithm introduced by Willems et al. 21] 5.1 Scheme: Prediction using CTW ....

....the next branch is taken . A prediction is made in the following way: P # w P # # w 8 : 5 1 # n # taken .5 1 # n # not taken otherwise # taken with probability #(P ) 6) where: #(P ) P # w P # # w . 5 1 p (n) 2 p (n) 7) It is not clear from [6] why a randomized predictor is used. Furthermore, it seems that a better way to determine the most likely branch outcome is to compare P # # w to P ## # w where P ## # w is the weight of the root node after assuming that the next branch was not taken . This solution may not be used however, ....

[Article contains additional citation context not shown here]

E. Federovsky, M. Feder, and S. Weiss. Branch prediction based on universal data compression algorithms. In Proceedings 25th Annual Symposium on Computer Architecture, pages 62--72, June 1998.

No context found.

Etian Federovsky, Meir Feder, and Shlomo Weiss. Branch prediction based on universal data compression algorithms. In Proceedings of the 25th Annual International Symposium on Computer Architecture, pages 62--72, June 1998.

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