Citation Context

...tion tracing. Based on the FACT framework, Zhai et al. employ a deterministic replay technique to predict the sequential computation time of one process in a parallel application on a target platform =-=[25]-=-. The main idea is to use the information recorded in the trace to simulate the execution result of MPI calls when there is actually only one MPI process, and utilize the deterministic data replay to ...

Citation Context

... et al. compare a set of different schemes for automated machine-based performance learning and prediction [22]. Zhai, Chen, and Zheng extrapolate single-node performance to complex parallel machines =-=[40]-=-. Wu and Müller [37] extrapolate traces to larger process counts and can thus predict communication operations. Their extrapolation relies on a trace compression scheme that assumes regular communicat...

Citation Context

... 30] but often look only at abstract models of execution and do not provide realistic estimates on the speedup of the refactored program. Yet other tools examine the scalability of a parallel program =-=[10, 53]-=- but look only at an existing implementation and therefore do not provide insight into the fundamental scalability of a program. Furthermore, a vast majority of existing tools assume that there is alr...

Citation Context

...han to actually run the analyzed code. Most performance prediction tools have a slowdown greater than one, i.e.[22, 23]. This places our implementation in the group of rapid prediction tools, such as =-=[14, 24]-=- . C. Reducing the slowdown Knowing that an important metric for classifying performance prediction tools is the slowdown, we previously proposed an optimized block benchmarking technique presented in...

Citation Context

...c 2016 ACM 978-1-4503-4092-2/16/03. . . $15.00 DOI: http://dx.doi.org/10.1145/http://dx.doi.org/10.1145/2851141.2851159 applications in a simple way, without having to understand in detail the internals of the application or the machine it will run on. ESTIMA enables developers to visualize the scalability of their applications, as well as to discover bottlenecks that might not be evident during initial performance benchmarking. ESTIMA can be applied with little effort to any parallel in-memory application, in contrast to other approaches that heavily rely on application-specific information [4, 6, 22, 24, 25, 30, 44]. Instead, ESTIMA leverages stalled cycles to extrapolate the scalability of an application. These are cycles the application spends on non-useful work, such as waiting for a cache line to be fetched from memory or waiting on a busy lock. Contention for shared resources typically increases with the number of cores used by an application, resulting in an increase in stalled cycles that directly impact the application’s scalability. The application’s performance keeps improving as long as adding more cores increases the number of useful cycles. As soon as adding more cores mostly results in stal...

Citation Context

...5%. In contrast, with respect to the target architecture our approach utilizes the published execution times of the SPEC CPU and IMB benchmarks, resulting in an error of at most 15%. The PHANTOM tool =-=[15]-=- uses deterministic replay techniques to execute any process of a parallel application on a single node of the target system at real speed, hence measuring computation performance. This assumes that a...

Citation Context

...ed runtime. Chiang et al. suggested a test run before running to acquire more accurate estimated runtime [13]. Zhai et al. developed a performance prediction tool to assist an accurate estimated time =-=[14]-=-. Tang et al. got usable information from historical information [15]. In order to improve the estimate accuracy, Cynthia Bailey Lee et al. gave a detailed survey. However, performance prediction is a...

Citation Context

... consider the problem of finding the hardware saturation point. Other researchers have looked at using techniques such as deterministic replay coupled with clustering to select representative replays =-=[11]-=-. A trace of the application is extracted from a host machine and is then replayed locally on a single node. Interestingly, it is possible to use this technique to execute multiple replays on the same...