A. Agarwal and M. Huffman. Blocking: Exploiting Spatial Locality for Trace Compaction. In Proceedings of the 17th Annual International Symposium on Computer Architecture, ACM, 1990.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....one usually has some knowledge of the future uses of the trace. Lossy trace reduction techniques attempt to exploit such knowledge so that the trace size is reduced dramatically but enough information is maintained for the intended uses. The simplest lossy reduction technique is bloct;ing [AH90]. Blocking replaces references to individual addresses with references to larger blocks of address space, such as memory pages. Subsequent references to addresses within the same page can then be reduced to a single reference. This reduction does not affect the simulation of time independent ....

....alternative to statistical trace synthesis techniques (e.g. Bab81] Finally, our techniques are complementary to lossy reduction algorithms that exploit different principles. Since the output of either of our algorithms is itself a trace, other trace reduction techniques can be applied (e.g. [JH94, AH90]) Furthermore, lossless techniques, including simple file compressors like gzip, can be applied to our reduced traces to yield much smaller files. 1They also note that, with one additional integer for the reduced trace, mean memory sizes can be accurately calculated for WS and VMIN. 3 Safely ....

A. Agarwal and M. Huffman. Blocking: Exploiting spatial locality for trace compaction. In _Proceedings, ACM SIGMETRICS, pages 48-57, 1990.

....cache configurations 1, finding an op timal cache configuration in terms of system performance metrics such as runtime (best, average, or worst case) power consumption, etc, requires fast and accurate performance estimation. In applying traditional trace driven cache simulation methods [4] 5] 6][7] to the performance estimation of multiple cache IPbased systems, we face a significant problem. In such methods, address traces are initially obtained and assumed invariant over various cache configurations. However, in multiple cache IP based systems, system behavior, i.e. address traces can ....

A. Agarwal and M. Huffman, "Blocking: Exploiting Spatial Locality for Trace Compaction", Proc. 1990.

....the reduced trace, and should be understood as well. One approach to trace reduction is to apply standard data compression algorithms. For example, the UNIX compress utility, which implements the Lempel Ziv algorithm [Ziv76] achieves a compression factor of about 3 to 5 on typical address traces [Agarwal90]. Samples has shown that much higher degrees of compression can be attained if the full address trace is first preprocessed to produce a stream of address differences [Samples89] When the resulting difference trace is input to the same Lempel Ziv compression algorithm, the compression factors ....

.... 4 5 Fully associative Memories Snapshot Method [Smith77] 5 100 0 4 50 No 4 5 Fully associative Memories Cache Filter [Puzak85] 10 20 0 Yes N A Fixed line size Caches [Wang90] 10 20 0 7 15 Yes N A Fixed line size Caches Block Filter [Agarwal90] 50 100 0 No 12 Fixed line size Caches Time Sampling [Laha88] 5 20 0 5 20 No 5 Small Caches ( 128 K byte) Kessler91] 10 0 10 No 10 Small Caches ( 1 M byte) Set Sampling [Puzak85] 5 10 0 10 No 2 Set Sample Not General ....

[Article contains additional citation context not shown here]

Agarwal, A. and Huffman, M. Blocking: Exploiting spatial locality for trace compaction. In Proceedings of the

....one usually has some knowledge of the future uses of the trace. Lossy trace reduction techniques attempt to exploit such knowledge so that the trace size is reduced dramatically but enough information is maintained for the intended uses. The simplest lossy reduction technique is blocking [AH90] Blocking replaces references to individual addresses with references to memory pages. Subsequent references to addresses within the same page can then be reduced to a single reference. This reduction does not affect the simulation of time independent paging algorithms algorithms that do not ....

....trace synthesis techniques (e.g. Bab81] Finally, we should mention that our techniques are complementary to reduction algorithms that exploit different principles. Since the output of either of our algorithms is itself a trace, other trace reduction techniques can be applied (e.g. JH94, AH90] As we will see, simple file compression of our reduced traces with the gzip utility yields much smaller files, further decreasing storage requirements. 74 3.3 The Algorithms 3.3.1 Safely Allowed Drop (SAD) Full traces commonly contain a large number of references that are ignored by ....

A. Agarwal and M. Huffman. Blocking: Exploiting spatial locality for trace compaction. In Proceedings, ACM SIGMETRICS, pages 48--57, 1990.

....to reduce a preexisting trace. Puzak [41] suggested an alternative based on the fact that references across sets tend to be highly correlated [44] select only a small number of the sets. Experience with reasonable data indicates that retaining only 10 percent of the sets is adequate. Blocking [45] serves as another alternative by taking the misses from Puzak s cache filter and passing them through a block filter. The block filter considers each window of w consecutive references and sends out a reference from each spatial locality contained within the window. 2.4 Use From a technique ....

A. Agarwal and M. Huffman, "Blocking: Exploiting spatial locality for trace compaction," in Proceedings of the 1990 ACM Sigmetrics Conference on Measurement and Modeling of Computer Systems, (Boulder, CO), pp. 48--57, May 1990.

....invisible to the operating system. The log buffer holds about 5.9 million eight byte log entries, which is enough for 5 20 seconds of real time. There is so much information in a single reconstructed trace that we have not been motivated to try stitching multiple traces together [CHRG95, AH90] a single reconstructed trace contains about 650 MB of dynamic i stream with instruction and data addresses. Recording the log in main memory is much faster than recording on disk or tape. Recording in physical memory instead of virtual memory allows us to trace the lowest levels of the ....

Anant Agarwal and M. Huffman. Blocking: Exploiting spatial locality for trace compaction. Performance Evaluation Review, 18(1):48--57, May 1990.

....the volume reduction of performance data taken from a restricted problem domain. For example, Wang and Baer [11] detail a reduction technique for memory reference traces that retains sufficient information to exactly reproduce the cache simulation obtained with the full trace. Argawal and Huffman [2] describe a similar method for compacting memory reference traces. Ball and Larus [4] present two algorithms which determine program profiling and tracing instrumentation points which significantly reduce the volume of performance data generated. Their approach, however, requires a computationally ....

Argawal, A., and Huffman, M. Blocking: Exploiting Spatial Locality for Trace Compaction. In Proceedings of the ACM SIGMETRICS Conference on the Measurement and Modeling of Computer Systems (May 1990), Association for Computing Machinery, pp. 48--57.

....Working Set [Denn68] Blocking is so widely applicable that it is practically assumed in most simulation work. For the remainder of this paper, when we refer to an original trace, we are referring to a blocked trace. Recent work on trace reduction includes the technique of Agarwal and Huffman [AgHu90]. Whereas most lossy reduction techniques concentrate on the temporal locality of a program trace, their approach exploits spatial locality and results in an extra significant factor of reduction. Other trace reduction methods include trace sampling and trace stripping (e.g. see [Puza85] Both ....

....to statistical trace synthesis techniques (e.g. Baba81] Finally, we should mention that our techniques are complementary to reduction algorithms that exploit different principles. Since the output of our algorithms is itself a trace, other trace reduction techniques can be applied (e.g. [JoHa94, AgHu90]) As we will see, simple file compression of our reduced traces with the gzip utility yields much smaller files, further decreasing storage requirements. 3 The Algorithms 3.1 Safely Allowed Drop (SAD) Full traces commonly contain a large number of references that are ignored by virtual memory ....

A. Agarwal and M. Huffman, "Blocking: Exploiting spatial locality for trace compaction ", Proc. SIGMETRICS '90, pp.4857.

....length of the trace; reducing the number of references in the trace has a significant impact on both storage required and simulation time. Previous methods exploited temporal locality for compressing traces and could yield up to an order of magnitude reduction in the length of the address trace [1]. The key idea behind trace compaction is that not all of the memory references carry the same amount of useful information. Traces can be reduced in size by retaining only those memory references that contain new information. The general approach is to find characteristics of the trace data that ....

.... data less useful for applications that do not match these assumptions [7] Thi d d i h F M k 4 0 4 r 23 Several methods have been proposed to reduce or compress address traces, including trace deletion and the snapshot method [8] trace stripping [6, 9] sampling [5] Mache [7] and blocking [1]. In the method called trace stripping, a direct mapped cache with a fixed block size is simulated. The misses and the run lengths of the hits are recorded from the simulation run, resulting in compaction on the order of 90 95 (a compression ratio of 10 to 20) Puzak proved that no error is ....

[Article contains additional citation context not shown here]

Agarwal, A. and M. Huffman, "Blocking: Exploiting Spatial Locality for Trace Compaction," Proceedings, ACM SIGMETRICS `90: 1990.

....format [3] and each simulation that reads this trace will consume several hours of CPU time on typical workstations. Seeking to reduce the costs in storage and simulation time associated with large traces, researchers have pursued a number of trace filtering, sampling, and compression techniques [2 9]. The key concept behind these reduction and compression methods is that much of the information stored in traces is redundant. Such redundant information can be removed without a severe impact on the accuracy of simulations as compared to results using the original traces. Typically, these ....

A. Agarwal and M. Huffman, "Blocking: Exploiting Spatial Locality for Trace Compaction," Proceedings of SIGMETRICS `90, pp. 48-57, ACM, 1990.

....and spatial chains of memory accesses. If it finds a chain that is dense enough, the generated code will not append new references to the trace for the redundant accesses. The compaction introduces an error in the memory simulation. The method is similar to those run time methods described in [5], 6] 7] which report consistently with our experience that the error is usually small. II. Redundancy in Memory Accesses DBE is designed for simulation of processor caches and disk buffers by using a memory access trace. The memory access trace compaction done by DBE is based on the notion ....

A. Agarwal and M. Huffman, "Blocking: Exploiting Spatial Locality for Trace Compaction," in Proceedings of 90' ACM SIGMETRICS, 1990, number 18(1) in Performance Evaluation Review, pp. 48--57.

....direct mapped caches need no cache replacement policy. This enables their use as trace lters without causing any error in simulating caches with a larger set size. The stack deletion and trace stripping methods are based on temporal analysis of a trace. Agarwal and Huoeman have presented a method [2], which applies both spatial and temporal analysis. Agarwal and Huoeman call the phases of their method temporal ltering and spatial blocking. The method presented in this thesis resembles the methods above. The major dioeerence between these methods and the method presented in this thesis is the ....

....cause errors in simulations [21] 3.4 Spatial blocking The previous compression methods are based solely on temporal analysis of a trace. Agarwal and Huoeman have presented a method, which applies also spatial analysis. Agarwal and Huoeman call their spatial trace compaction spatial blocking [2]. Spatial blocking is designed to be done to traces, which are already temporally ltered. The temporal ltering, that Agarwal and Huoeman use, is based on trace stripping. The spatial blocking method is based on statistics. It assumes that all the memory references in a spatial neighborhood have ....

[Article contains additional citation context not shown here]

A. Agarwal and M. Huoeman. Blocking: Exploiting Spatial Locality for Trace Compaction. In Proceedings of 90' ACM SIGMETRICS, number 18(1) in Performance Evaluation Review, pages 4857, 1990.

.... prefetching) or interactions between sets (e.g. a single write buffer) We do not consider other (non sampling) techniques that reduce trace data storage, such as, Mache [SAMP89] stack deletion and snapshot method [SMIT77] trace (tape) stripping [PUZA85, WANB90] or exploiting spatial locality [AGAH90]. These techniques can be used in addition to the sampling considered in this study. We also do not consider Przybylski s prefix technique [PRZY88] which prepends all previously referenced unique addresses to each time observation. This method seems unattractive for multi megabyte caches where ....

A. AGARWAL and M. HUFFMAN, "Blocking: Exploiting Spatial Locality for Trace Compaction," Proceedings of the Conference on Measurement and Modeling of Computer Systems, 1990, pp. 48-57.

....gains translate only into storage savings not simulation speedup. In general, since the output of olr is itself a trace, all standard trace reduction techniques can be used to reduce it further. Such techniques could, for instance, exploit the spatial locality of the reference trace (e.g. [AgHu90]) and result in an extra significant factor of reduction. 4 Related Work There are numerous pieces of work on various aspects of the LRU replacement policy and LRU based data structures. Indicatively, we mention Sleator and Tarjan s well known results [SlTa85] and some more recent treatment of ....

A. Agarwal and M. Huffman, "Blocking: Exploiting spatial locality for trace compaction", Proc. SIGMETRICS '90, pp.48-57.

....Working Set [Denn68] Blocking is so widely applicable that it is practically assumed in most simulation work. For the remainder of this paper, when we refer to an original trace, we are referring to a blocked trace. Recent work on trace reduction includes the technique of Agarwal and Huffman [AgHu90]. Whereas most lossy reduction techniques concentrate on the temporal locality of a program trace, their approach exploits spatial locality and results in an extra significant factor of reduction. Other trace reduction methods include trace sampling and trace stripping (e.g. see [Puza85] Both ....

....to statistical trace synthesis techniques (e.g. Baba81] Finally, we should mention that our techniques are complementary to reduction algorithms that exploit different principles. Since the output of our algorithms is itself a trace, other trace reduction techniques can be applied (e.g. [JoHa94, AgHu90]) As we will see, simple file compression of our reduced traces with the gzip utility yields much smaller files, further decreasing storage requirements. 3 The Algorithms 3.1 Safely Allowed Drop (SAD) Full traces commonly contain a large number of references that are ignored by virtual memory ....

A. Agarwal and M. Huffman, "Blocking: Exploiting spatial locality for trace compaction", Proc. SIGMETRICS '90, pp.48-57.

....protocols and other system design issues. However, this technique requires a parallel trace for multiprocessor systems which is difficult and expensive to obtain. Although different techniques in gathering program traces and in reducing the complexity of tracedriven simulation have been developed [1, 2, 3, 4], the complexity of this technique increases as the system size increases due to the size of address traces and synchronization requirements. Most importantly, an address trace can only capture a tiny fraction of program behavior. Therefore, whether the trace can represent a typical workload is ....

A. Agarwal and M. Huffman, "Blocking: Exploiting spatial locality for trace compaction," in Proc. 1990 ACM SIGMETRICS Conf. Meas. Comput. Syst., pp. 48--57, 1990.

....requirements for address traces. According to Larus [113] a 10 MIPS processor produces on average 1 A similar technique has been used in the DDM emulator, albeit not for address tracing. 7.1 Evaluation techniques 77 40 60 Mbytes of trace data in one second of real execution. Some schemes [6] have been proposed that deal with compaction of trace data. To overcome the space overhead problem, execution driven simulators are becoming commonplace. 7.1.2 Execution driven simulation Many researchers have criticised the validity of trace driven simulation techniques [30, 51] Bitar argues ....

A. Agarwal and M. Huffman. Blocking: Exploiting Spatial Locality for Trace Compaction. In Proceedings of the 17th Annual International Symposium on Computer Architecture, ACM, 1990.

....therefore invisible to the operating system. The log buffer holds about 5.9 million eightbyte log entries, which is enough for 5 20 seconds of real time. There is so much information in a single reconstructed trace that we have not been motivated to try stitching multiple traces together [CHRG95, AH90] a single reconstructed trace contains about 650 MB of dynamic i stream with instruction and data addresses. Recording the log in main memory is much faster than recording on disk or tape. Recording in physical memory instead of virtual memory allows us to trace the lowest levels of the ....

Anant Agarwal and M. Huffman. Blocking: Exploiting spatial locality for trace compaction. Performance Evaluation Review, 18(1):48--57, May 1990.

....a synthetic trace generator. 5 Workload Model Ideally, a multiprocessor trace would feeds references into this simulation framework. However, this is difficult to obtain and does not provide the required flexibility. Although different techniques in gathering parallel traces have been developed [15, 16, 17], the complexity of trace driven simulation increases as the system size increases. A reasonable alternative is to generate artificial traces with approximately the same characteristics that real traces might have. Since the system to be evaluated usually does not exist, the synthetic trace must ....

A. Agarwal and M. Huffman, "Blocking: Exploiting spatial locality for trace compaction," in Proc. 1990ACM SIGMETRICS Conf. Meas. Comput. Syst., pp. 48--57, 1990.

....trace. The results in [14] show that as few as 35 samples can give accurate estimates of both the mean value and the distribution of the miss ratio. The advantage of this scheme is that with only a small fraction of the total address trace the original program behavior can be reproduced. In [1] and [24] alternative schemes are presented for trace compaction and efficient trace driven simulation. We chose the simpler approach because it was satisfactory for our purposes, saving both disk space and cache simulation time. For the results in this report we took 50 samples for each benchmark ....

A. Agarwal and M. Huffman. Blocking: Exploiting Spatial Locality for Trace Compaction. In Conference on Measurement and Modeling of Computer Systems, pages 48--57. ACM SIGMETRIC, May 1990.

....of 1.05 to 4. b) Snapshots which only record a set of references every T references. The range of error rates caused by the snapshot method is similar to the first method while producing a compression ratio of 5 to 100. These two schemes have been shown to be primarily suitable for paging studies [1]. Puzak described a trace stripping scheme that employs a small direct mapped filter cache to remove redundant references. This technique introduces no error in simulations of caches with the same line size as the filter cache [6] However, when the line size is varied, significant errors in miss ....

.... This technique introduces no error in simulations of caches with the same line size as the filter cache [6] However, when the line size is varied, significant errors in miss ratios are produced [3] Agarwal introduced a two step trace compaction scheme: a cache filter followed by a block filter [1]. Errors in simulated miss ratios ranged from 1 to 16 , while the compression ratio reached 15 under some configurations of the cache filter and block filter. In addition to the errors produced in cache simulations, another problem with these trace filtering techniques is that traces are used to ....

Agarwal, A. and M. Huffman, "Blocking: Exploiting Spatial Locality for Trace Compaction," Proceedings, ACM SIGMETRICS 1990.

.... prefetching) or interactions between sets (e.g. a single write buffer) We do not consider other (non sampling) techniques that reduce trace data storage, such as, Mache [SAMP89] stack deletion and snapshot method [SMIT77] trace (tape) stripping [PUZA85, WANB90] or exploiting spatial locality [AGAH90]. These techniques can be used in addition to the sampling considered in this study. We also do not consider Przybylski s prefix technique [PRZY88] which prepends all previously referenced unique addresses to each time observation. This method seems unattractive for multi megabyte caches where ....

A. AGARWAL and M. HUFFMAN, "Blocking: Exploiting Spatial Locality for Trace Compaction," Proceedings of the Conference on Measurement and Modeling of Computer Systems, 1990, pp. 48-57.

.... Filter [Smith77] 5 100 0 4 50 No 4 5 Fully associative Memories Snapshot Filter [Smith77] 5 100 0 4 50 No 4 5 Fully associative Memories Cache Filter [Puzak85] 10 20 0 Yes N A Fixed line size Caches [Wang90] 10 20 0 7 15 Yes N A Fixed line size Caches Block Filter [Agarwal90] 50 100 0 No 12 Fixed line size Caches Time Sampling [Laha88] 5 20 0 5 20 No 5 Small Caches ( 128 K byte) Kessler91] 10 0 10 No 10 Small Caches ( 1 M byte) Set Sampling [Puzak85] 5 10 0 10 No 2 Set Sample Not General [Kessler91] 10 0 10 No 10 Constant bits Set ....

....use of the reduced trace. 5.1 Trace Compression One approach to trace reduction is to apply standard data compression algorithms. As an example, the UNIX compress utility, which implements the Lempel Ziv algorithm [Ziv76] achieves a compression factor of about 3 to 5 on typical address traces [Agarwal90]. Samples showed that much higher degrees of compression can be attained if a full address trace is first preprocessed to produce a difference trace, as is done in Mache [Samples89] Mache computes a difference trace by dividing a full address trace into substreams according to some separation ....

[Article contains additional citation context not shown here]

Agarwal, A. and Huffman, M. Blocking: Exploiting spatial locality for trace compaction. In Proceedings of the 1990 SIGMETRICS Conference on Measurement and Modeling of Computer Systems, Boulder, CO, ACM, 48-57, 1990.

....of massively large traces for virtually every conceivable workload, trace driven simulation has become the dominant method for cache performance analysis. Unfortunately, simulations are not only slow, but they lend little insight into the behavior of caches. Statistical methods for trace analysis [1, 2, 3] help reduce the simulation time, but they are still significantly slower than modeling, and offer little additional insight over full simulation. Because caches are amenable to mathematical analysis, we contend that simulations against massively long traces are unnecessary, except in the final ....

Anant Agarwal and Minor Huffman. Blocking: Exploiting Spatial Locality for Trace Compaction. In Proceedings of ACM SIGMETRICS 1990.

No context found.

A. Agarwal and M. Huffman. Blocking: Exploiting Spatial Locality for Trace Compaction. In Proceedings of the 17th Annual International Symposium on Computer Architecture, ACM, 1990.

No context found.

A. Agarwal and M. Hu#man, "Blocking: Exploiting spatial locality for trace compaction", Proc. SIGMETRICS '90, pp.48-57.

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