F. Gabbay and A. Mendelson. (1997). Can Program Profiling Support Value Prediction? In Proceedings of the 30th Micro, pp. 270-280  Home/Search   Document Details and Download   Summary   Related Articles   Check

....of the Next Info list that corresponds to the value with the maximum counter value. In Figure 3, for example, since the maximum counter value in the extracted PHT entry is larger than the predefined threshold, and since Age is two, the index of the value corresponding to the maximum counter, [10], will be the starting point of the search through the Next Info list. We need to follow the Next Info list Age = 2 times, i.e. Next Info[10] 11] Next Info[11] 00] The index [00] then is the index for the predicted value. However, since the value of Age can be as large as eight in this ....

....value in the extracted PHT entry is larger than the predefined threshold, and since Age is two, the index of the value corresponding to the maximum counter, 10] will be the starting point of the search through the Next Info list. We need to follow the Next Info list Age = 2 times, i.e. Next Info[10] = 11] Next Info[11] 00] The index [00] then is the index for the predicted value. However, since the value of Age can be as large as eight in this study, up to eight links may need to be followed, which could introduce a long delay to make a prediction. In order to reduce this delay, we add ....

F. Gabbay and A. Mendelson. "Can Program Profiling Support Value Prediction?" 30 Symposium on Microarchitecture (MICRO), pages 270-280, December 1997.

....Because making no prediction and waiting for the memory access to complete is faster than making an incorrect prediction and having to recover from it, most predictors in the literature include a confidence estimator. Confidence estimators inhibit predictions that are likely to be incorrect [7, 9, 14, 15, 17, 18, 23] and thus reduce the number of mispredictions and the associated recovery cost, which improves the predictor s overall performance. The bimodal confidence estimator [14, 15, 17] is based on saturating up down counters and can be described using four parameters: a maximum, a threshold, a penalty ....

F. Gabbay, A. Mendelson. Can Program Profiling Support Value Prediction? 30 Annual ACM/IEEE International Symposium on Microarchitecture, 1997, pages 270-280.

....requires predicting more than one bit of information. Skepticism regarding these techniques is often centered on the difficulty of predicting 32 (or 64) bits of information. Locality in the value history of a program indicates that the information content is significantly less than 32 bits worth [6, 8, 17, 18, 25]. Exploiting this locality requires a significant investment in hardware for modest prediction accuracy [7, 18, 25, 30] But if the hardware cost can be tolerated and high accuracy can be reached, value prediction can reduce pipeline stalls by introducing more (speculatively) independent ....

....value locality was broadly classified into two categories: computational and context based [25] Most of the proposed value predictors exploit these localities in the local value history. Those predictors include the last value predictor [18] the last N value predictor [4] the stride predictor [7, 8, 17, 18], and context predictors [9, 25, 30] The hybrid predictors can combine both computational and context based predictors to exploit both types of localities to achieve higher prediction capability [21, 22, 25, 30] Compared to local value history, using global value history for prediction is less ....

F. Gabbay and A. Mendelson, "Can program profiling support value prediction?", in 30 International Symposium on Microarchitecture (MICRO-30), Nov. 1997.

....some cases, an order of magnitude more area. The other key difference is that the above approaches also target non arithmetic instructions and use dynamic replacement, while our current version does not require these additions to produce comparable or better performance. Gabbay and Mendelson in [6,7] presented an approach that used profiling to enhance the performance of value prediction. The value prediction hardware only predicts the values for instructions that have good value locality and a high probability of a correct prediction, as determined by the profiler. While this approach ....

F. Gabbay and A. Mendelson; "Can Program Profiling Support Value Prediction"; International Symposium on Microarchitecture, 1997

....locality than data loads, instruction address loads hold an edge over data address loads, and integer data values are more predictable than floating point data values. Our approach separates loads into many more classes that exhibit greatly varying predictability behavior. Gabbay and Mendelson [15] explore the possibility of using program profiles to enhance the e#ciency of value prediction. They use profiling to insert opcode directives to filter out highly unpredictable values from being allocated in the load value predictor, which considerably reduces the amount of aliasing. Our approach ....

F. Gabbay and A. Mendelson. Can program profiling support value prediction? In 30th Annual ACM/IEEE International Symposium on Microarchitecture, 1997.

....the fixed width nature of the word model is a mismatch for certain computations, e.g. boolean logic decisions and arithmetic with small numbers. In such cases, a substantial number of bits in a word contribute no useful information to the computation. In addition, there is evidence [42] 45] [13] [6] that a significant fraction of word operations on traditional word based architectures are continually repeated with previously seen inputs, contributing no useful work in steady state. These observations suggest that the traditional word model, despite its convenience, lends itself to ....

....waste, and consequently to degraded run time, area, and or power in any implementation. Existing architectures have sought to overcome wasteful behavior of the word model in several ways. In microprocessors, wasteful word operations can be reduced by value prediction hardware [32] 42] [13] [45] as well as by value specific optimizations (VSO) using run time partial evaluation and specialization [26] 10] 1 [30] 19] Wasteful bit operations can be reduced in part by going to finer grained architectures. Recent multimedia instruction sets (e.g. Vis [28] MMX [36] AltiVec [35] ....

[Article contains additional citation context not shown here]

Freddy Gabbay and Avi Mendelson. Can program profiling support value prediction? In Proc. 30th International Symposium on Microarchitecture (MICRO-30), pages 270--280, Research Triangle Park, North Carolina, December 1--3, 1997.

....prediction, however, is yet a very novel technique, whose actual efficacy has to be still proved. Recent works on value prediction have shown that data are predictable [Lipa96a] Saze97a] Wang97] and several techniques for value prediction have been proposed [Lipa96a] Lipa97] Saze97a] [Gabb97], Nakr99] The results obtained are very promising: using sophisticated value predictors, we can predict correctly about 60 of the output values in a program. However, only a few works have shown the actual potential of value prediction to increase ILP under real conditions [Rych98] ....

....context based predictors. Most of the value predictors proposed in the literature fit in one of the following types: x Last value predictors (LVP) that make a prediction based on the last outcome of the same static instruction, and can predict correctly constant sequences of data. Lipa96a] [Gabb97], Wang97] x Stride predictors (SP) that make a prediction based on the last outcome plus a constant stride, and can predict correctly arithmetic sequences of data (even constant sequences, whose stride is 0) Gabb97] Nakr99] Wang97] x Context based predictors (CBP) which learn the ....

[Article contains additional citation context not shown here]

F. Gabbay and A. Mendelson, "Can Program Profiling Support Value Prediction?", Proc. of the 30 th Int. Symp. on Microarchitecture (MICRO-30), pp. 270-280, Dec. 1997.

....programmer specified variables exhibit predictable behavior at runtime. Our implementation automates this process, identifying candidates for profiling, and then choosing among these candidates for optimization. There is a large body of work relating to value prediction and speculation, e.g. see [11, 12] . This involves using a combination of compiler generated information and runtime information. For example, a load instruction will often load the same value as it loaded the previous time it was executed. Predictable instructions such as these can be speculatively executed, and then checked at a ....

F. Gabbay and A. Mendelson. Can program profiling support value prediction? In Proceedings of the 30 th Annual ACM/IEEE International Symposium on Microarchitecture, pages 270--280, Dec 1997.

....predictor is an improved version of the conventional stride predictor. It maintains two separate strides. The stride used for making predictions is only updated if a new stride has been seen twice in a row, which significantly reduces the number of mispredictions. Profiling: Gabbay and Mendelson [7] explore the possibility of using program profiling to enhance the efficiency of value prediction. They use their profiling results to insert opcode directives that allow them to allocate only highly predictable values, which reduces the amount of aliasing. However, even manual fine tuning of the ....

....of whether the branches were taken or not. We use the same technique. Summary: The novelty of our predictor is that it uses prediction outcome histories, an idea taken from the branch prediction literature, for confidence estimation instead of saturating counters. While others have used profiling [2, 7], our use of profiling is different in so far that we do not profile actual load values and do not need to modify the binaries in any way. 5. Methodology All our measurements are performed on the DEC Alpha AXP architecture [3] To perform a thorough design space evaluation, we instrumented our ....

F. Gabbay, A. Mendelson. "Can Program Profiling Support Value Prediction?". 30 th International Symposium on Microarchitecture. December 1997.

....value locality than data loads, instruction address loads hold an edge over data address loads, and integer data values are more predictable than floating point data values. Our approach is able to separate loads into many more classes with quite different predictabilities. Gabbay and Mendelson [10] explore the possibility of using program profiles to enhance the efficiency of value prediction. They use profiling to insert opcode directives to filter out highly unpredictable values from being allocated in the load value predictor, which considerably reduces the amount of aliasing. However, ....

F. Gabbay and A. Mendelson. Can program profiling support value prediction? In 30th Annual ACM/IEEE International Symposium on Microarchitecture, 1997.

....in the predictor for both predictable and unpredictable loads. Hence, it may be advantageous to combine a static or profile based approach with a dynamic predictor to filter out the unpredictable loads so that the predictor only has to be designed large enough to handle the predictable loads [GaMe97]. Another important advantage of the dynamic approach is that it can adapt to changes in the program behavior during the course of the execution. The information provided by static approaches or by profiles is normally fixed and cannot be changed at runtime. Due to the limitations of the static ....

....further show that a stride predictor can also be made storage less in combination with a last two value predictor (Chapter 9) Most of the above predictors and many more are discussed and compared performance wise in Section 5.4 and Chapter 8. 10.3 Profile based Approaches Gabbay and Mendelson [GaMe97] explore the possibility of using program profiles to enhance the efficiency of value prediction. They use profiling to insert opcode directives to filter out highly unpredictable values from being allocated in the load value predictor, which considerably reduces the amount of aliasing. However, ....

F. Gabbay, A. Mendelson. "Can Program Profiling Support Value Prediction ?". 30 th Annual ACM/IEEE International Symposium on Microarchitecture. December 1997.

....ASPLOS IX, Cambridge, MA. Copyright 2000 ACM 1 58113 199 2 00 0006 . 5.00 1. INTRODUCTION Recent research has demonstrated that values produced by executing instructions exhibit a high degree of value locality, that is, multiple executions of the same instruction often produce the same value [3, 8]. Value locality has been exploited in the design of value reuse and prediction mechanisms for superscalar processors. In this paper we show that another kind of locality, that we refer to as the frequent value locality, is also quite prevalent in programs. If we track the values involved in ....

....1 3 7 099.go 18.8 0 0.07 0.5 124.m88ksim 4.1 0 63 70 126.gcc 0.3 0 10 18 130.li 1.2 0 0.3 0.3 134.perl 20.9 0 0.3 0.4 147.vortex 8.2 0 9 29 Table 3: Finding frequently accessed values. Frequent value locality vs value locality. The traditional notion of value locality [3, 8] is quite distinct from frequent value locality. In particular, value locality is defined in context of individual instructions and is observed by examining the values generated by multiple executions of a specific instruction. Therefore it is not surprising that value locality has found ....

F. Gabbay and A. Mendelson, "Can Program Profiling Support Value Prediction?, " Proc. of the 30th ACM/IEEE International Symp. on Microarchitecture, pages 270-280, Dec. 1997.

....behavior (e.g. value prediction [12, 29, 35, 38] and dynamic specialization [3, 15, 19] the identification of candidate invariant values is either a strict requirement or an enhancement. Because requiring the programmer to identify these values is tedious and error prone, value profiling [9, 10, 11, 20] has been proposed to automatically characterize value invariance (the frequency at which the value is seen) as well as to identify the most common values. Early implementations of value profiling using instrumentation or simulation are not feasible for large programs (Calder, et al. reports an ....

F. Gabbay and A. Mendelson. Can program profiling support value prediction? In Proc. 27th International Symposium on Microarchitecture, pages 270--280, Dec. 1997.

....the same result value computed during their last execution. Several subsequent proposals have been made for improved hardware value predictors [11, 15, 14] Gabbay and Mendelson explored the use of profiling techniques to identify instructions which exhibit a high degree of value locality [10]. They showed that hardware value misprediction rates could be reduced by tagging the opcodes of predictable instructions, marking them as candidates for hardware prediction. Our low overhead value sampling techniques could be used to provide even more detailed information to such hardware ....

F. Gabbay and A. Mendelson. "Can program profiling support value prediction?". In Proceedings of the 30th International Symposium on Microarchitecture, pages 270--280, Dec. 1997. URL: http://www-ee.technion.ac.il/fredg/micro30.ps.gz or http://www.acm.org/pubs/articles/proceedings/micro/266800/p270gabbay /p270-gabbay.pdf.

....Methods is described and termed context prediction. Two level correlated structures [29] 31] are able to capture more complicated repeating patterns. 22] proposes efficient hybrid predictors to reduce the cost of context prediction. Software profiling for value prediction has been done in [3][7]. Other works apply value prediction to specific microarchitectural tasks. Numerous works use value prediction to predict load values or effective addresses [4] 6] 9] 11] 12] 16] 18] 1] 5] 15] 23] predict inter thread data dependencies in multithreaded architectures. 10] studies the ....

F. Gabbay and A. Mendelson, "Can Program Profiling Support Value Prediction?," in Micro-30, pp. 270-280, December 1997

....execution . In order to maintain the correctness of the execution, a speculatively executed instruction can only retire if the prediction it relies upon was proven to be correct; otherwise, it is discarded. In this paper we study the concept of value prediction introduced in [25] 26] 16] [17] and [18] The new approach attempts to eliminate true data dependencies by predicting at run time the outcome values of instructions, and executing the true data dependent instructions based on that prediction. Moreover, it has been shown that the bound of true data dependencies can be exceeded ....

.... value prediction aims at allowing the processor to execute operations beyond the limit of true data dependencies (given by the dataflow graph) Related work: The pioneer studies that introduced the concept of value prediction were made by Lipasti et al. and Gabbay et al. 25] 26] 16] [17] and [18] Lipasti et al. first introduced the notion of value locality the likelihood of a previously seen value to repeat itself within a storage location. They found that load instructions tend to exhibit value locality and they suggested exploiting this property in order to reduce memory ....

[Article contains additional citation context not shown here]

F. Gabbay and A. Mendelson. Can Program Profiling Support Value Prediction? Proceedings of the 30 th Annual ACM/IEEE International Symposium on Microarchitecture, December, 1997.

No context found.

F. Gabbay and A. Mendelson. (1997). Can Program Profiling Support Value Prediction? In Proceedings of the 30th Micro, pp. 270-280

No context found.

F. Gabbay and A. Mendelson. (1997). Can Program Profiling Support Value Prediction? In Proceedings of the 30th International Symposium on Microarchitecture, pp. 270-280

No context found.

F. Gabbay and A. Mendelson. (1997). Can Program Profiling Support Value Prediction? In Proceedings of the 30th International Symposium on Microarchitecture, pp. 270-280

No context found.

F. Gabbay and A. Mendelson, "Can program profiling support value prediction ?," in Proc. Int. Symp. Microarchitecture, 1997, pp. 270--280.

No context found.

F. Gabbay and A. Mendelson. Can program profiling support value prediction? In Proceedings of the 30 th Annual ACM/IEEE International Symposium on Microarchitecture, pages 270--280, Dec 1997.

No context found.

Freddy Gabbay and Avi Mendelson. Can Program Profiling Support Value Prediction? In Proceedings of the 30th International Symposium on Microarchitecture, pages 270--280, Research Triangle Park, NC, USA, December 1997.

No context found.

F. Gabbay, A. Mendelson. "Can Program Profiling Support Value Prediction?" 30 on Microarchitecture. December 1997.

No context found.

F. Gabbay and A. Mendelson. (1997). Can Program Profiling Support Value Prediction? In Proceedings of the 30th International Symposium on Microarchitecture, pp. 270-280.

No context found.

F. Gabbay and A. Mendelson. Can program profiling support value prediction? In Proceedings of the 30 th Annual ACM/IEEE International Symposium on Microarchitecture, pages 270--280, Dec 1997.

First 50 documents

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