Citation Context

...lopment of performance equations – which were the main component of our strategy – is not well understood and cannot be automated at this time. There are, however, some promising results in this area =-=[3]-=-. A second perhaps equally important conclusion of this study is that there is still a significant gap in performance between the code generated by ATLAS and the BLAS routines. Although we do not unde...

Citation Context

...representative samples [15]. Static analysis is a completely different approach. It is based on one of the more well-known cache miss equations, such as the ones suggested by Martonosi [9] and others =-=[7, 18]-=-. Static analysis is a potentially very fast method. Static analysis can also be parameterized in terms of input data size etc. However, statistical analysis still requires A B C A C D B E Time d = 2 ...

Citation Context

...ly from two directions: program analysis and tracedriven cache simulation. Dependence analysis analyzes data reuses in loop nests and can estimate the number of capacity misses in scientific programs =-=[16]-=-, [17], [18], [19]. Other researchers used various types of array sections to measure data access in loops and procedures [20], [21], [22], [23], [24]. One limitation of dependence analysis is that it...

Citation Context

... is an estimate for the working set contribution of each memory operation. For this work, we leverage previous research for working set estimates based on reuse distances of load and store operations =-=[2, 21]-=-. Algorithm 1 illustrates the calculation of this estimate. Input into the algorithm is a list of cache block references for every load and store operations8 5 5 5 12 3 13 15 PE 1 PE 2 11 5 6 10 11 5 ...

Citation Context

...of interacting processors, caches, memories, I/O devices, and custom hardware, all of which can be customized to the application. Cache structures are especially difficult to simulate at a high level =-=[3]-=- [11], with few current tools to accurately simulate cache structures above the instruction set level. We tackle this problem by developing a method for simulating cache structures at a high level qui...

Citation Context

...table. Using hardware performance counters, we compare the actual number of cache misses per tile, against the expected number of cache misses, which is calculated analytically, using stack distances =-=[2]-=-.If the measured number of cache misses exceeds the expected number by a threshold, we assume interference due to cache sharing. The threshold is allowed to compensate for potential self and cross-int...

Citation Context

...hong found predictable patterns in the overall locality but did not consider the phase behavior [11]. For regular programs, the reuse distance can be directly measured by a compiler without profiling =-=[5]-=-. We are not aware of any trace-based technique that identifies static phases through locality analysis on program traces. Recently, a number of groups have studied interval phases. Duesterwald et al....

Citation Context

...reduces the chance of a cache miss. At the program level, reuse distance analysis extends dependence analysis, which identifies reuses of program data [1], to count the volume of the intervening data =-=[4, 8, 10]-=-. At the trace level, the analysis can correlate the change in locality in different runs to derive program-level patterns and complement static analysis [21, 30, 49]. To review the conversion formula...

Citation Context

.... To improve performance, we divide the priority list into partially sorted groups. For example, there are 4 windows at the 25th access in the example in Table 1: [25], [21, -22], [19, -22,- 23], and =-=[16,13,10, 0]-=- 1 . The swap loops are changed to iterate over the groups. The minimal element of a group is simply the last element. Grouping in priority lists was first invented by Sugumar and Abraham for simulati...