Smirni, E., and Reed, D. A., "Workload Characterization of Input/Output Intensive Parallel Applications," Proceedings of the Ninth International Conference on Modeling Techniques and Tools for Computer Performance Evaluation, June 1997  Home/Search   Document Details and Download   Summary   Related Articles   Check

....these issues, we are investigating the performance directed selection of le striping distributions across storage devices and redundant storage of multiple distributions to reduce access time. This exploration builds atop our earlier work on physical and logical, input output pattern comparisons [9, 10] and portable, parallel le systems [7] It integrates real time performance data, automatic access pattern classi cation, and fuzzy logic controls for choosing and con guring exible policies. The foundation of the research is a prototype software library called PPFS II (Portable Parallel File ....

....is NP complete. Input output characterization. There are a number of studies that present models of physical disk [23, 24, 25] and disk array [17] access behavior. Also, the logical and physical patterns of application input output in parallel scienti c applications have been studied extensively [9, 26, 8, 27, 10, 28]. These studies have shown that parallel applications exhibit a wide variety of input output request patterns. Insights from these studies led to new, parallel le system application programming interface (API) standardization e orts like the SIO API [29] and the MPI IO API [30] Flexible ....

E. Smirni and D. A. Reed, \Workload Characterization of Input/Output Intensive Parallel Applications, " in Proceedings of the 9th International Conference on Modelling Techniques and Tools for 13 Computer Performance Evaluation, pp. 169-180, Springer-Verlag, June 1997.

....that can estimate data access collisions. 4.3. Input Output For many high performance applications, the input output barrier rivals or exceeds that for computation, making design of scalable input output architectures and le systems a key aspect of any highperformance system. Our experience [7] has shown that software layering is a potential source of performance loss. Hence, we are developing an instrumented versions of multilevel input output libraries that enable multilevel analysis of I O requests (i.e. requests from the PACI ASCI Hierarchical Data Format (HDF 5) library, the ....

....Building on our input output characterizations, we are also developing queueing models of input output scalability and le striping. These models predict the performance of parallel disk systems as a function of request rates and access patterns, request sizes, and disk hardware parameters [7]. They can be used to estimate I O system scalability. 4.4. Networks and Computational Grids With increasing development of applications for heterogeneous, distributed computing grids, performance measurement and prediction must include application tuning for heterogeneous resources. The ....

E. Smirni and D. A. Reed. Workload Characterization of Input/Output Intensive Parallel Applications. In ########### ## ### ### ############# ########## ## ######### ########## ### ##### ### ######## ####### ##### ##########, June 1997.

....policies based on request patterns and system load, To address these issues, this thesis investigates the performance directed selection of file striping distributions across storage devices. This exploration builds atop our earlier work on physical and logical input output pattern comparisons [12, 13] and portable, parallel file systems [10] It integrates real time performance data, automatic access pattern classification, and fuzzy logic con4 trols for choosing and configuring flexible policies. The foundation of the research is a prototype software library called PPFS II (Portable Parallel ....

....about the application access pattern can enable file system optimizations that can increase the performance by more than an order of magnitude. With this premise, the logical and physical patterns of application input output in parallel scientific applications have been studied extensively [11 13, 45 47]. These studies have shown that parallel applications exhibit a wide variety of input output request patterns, with both very small and very large request sizes, sequential and non sequential access, and a variety of temporal variations. Insights from these studies led to new, parallel filesystem ....

[Article contains additional citation context not shown here]

E. Smirni and D. A. Reed, "Workload Characterization of Input/Output Intensive Parallel Applications," in Proceedings of the 9th International Conference on Modelling Techniques and Tools for Computer Performance Evaluation, pp. 169--180, Springer-Verlag, June 1997.

....application I O streams. We also develop prediction strategies that use the model to create sequences of I O accesses that match the application request pattern with high accuracy. 1.2 Related Work Much work has been done to describe the access patterns of scienti c applications. Smirni and Reed [11] characterize the I O accesses of ve scienti c applications from the Scalable I O initiative. Those applications exhibited access patterns ranging from simple sequential accesses to interleaved patterns across processors. While simple descriptions are sucient to describe these patterns, the ....

Evgenia Smirni and Daniel A. Reed. Workload Characterization of Input/Output Intensive Parallel Applications. In Proceedings of the 9th International Conference on Modelling Techniques and Tools for Computer Performance Evaluation, pages 169-180. Springer-Verlag, June 1997.

....applications when the performance of the application consisted of communication and computation, but did not account for other execution behaviors. Two execution behaviors we plan to add to the structural modeling 231 232 approach are I O behavior, perhaps building off work by Smirni [SRar, RSSS98] and memory behavior, possibly using work presented in PMH [ACFS93] or the LDA Model [SW98a] as starting points. In addition, we would like to expand the classes of applications that structural models have been shown to apply to. Expanding the use of structural models. We demonstrated ....

Evgenia Smirni and Daniel A. Reed. workload characterization of input/output intensive parallel applications. Performance Evaluation, to appear.

....10 XP S. Application areas include modeling of electron molecule collisions, a 3 D numerical simulation of the Navier Stokes equations, an implementation of the Hartree Fock self consistent field method to calculate the electron density around a molecule, and quantum chemical reaction dynamics [18, 81, 82]. These applications compose a snapshot of current input output practices on scalable parallel machines and reflect the programmers input output design choices based on the capabilities of available input output systems. They are but a small part of the Scalable Input Output Initiative s (SIO) ....

Smirni, E., and Reed, D. A. Workload Characterization of Input/Output Intensive Parallel Applications. In Modelling Techniques and Tools for Computer Performance Evaluation (June 1997).

....ViPErHiSS (most notably, evaluation of designs) but with a focus on a different application, specifically, that of performance evaluation of software architectures . Software performance engineering has also been investigated in the context of client server systems [25] The Pablo analysis tool [27] and Paradyn tools [ have been used for extracting workload characteristics of parallel scientific applications. Uysal et al. [30] have characterized and modeled I O workloads of data intensive applications intended for running on large scale parallel machines for the purpose of their ....

Evgenia Smirni and Daniel A. Reed. Workload characterization of input/output intensive parallel applications. In Proceedings of the Conference on Modelling Techniques and Tools for Computer Performance Evaluation, Springer-Verlag Lecture Notes in Computer Science, pages 169--180, June 1997.

....by these studies are useful for designing I O systems of parallel machines. However, these studies were limited by the small number of applications each of them studied. In particular, none of them included non scientific applications. Finally, Cypher et al.[14] Reddy et al.[38] and Reed et al. [12, 43] analyzed the behavior of groups of parallel applications that perform I O. Cypher et al. determined the steady state I O rate for a set of eight programs. This helps answer one part of one of the questions mentioned above. Reddy et al. studied parallel versions of the Perfect Club benchmarks and ....

E. Smirni and D. A. Reed. Workload characterization of input/output intensive parallel applications. In Modelling Techniques and Tools for Computer Performance Evaluation, June 1997. to appear.

....grids, we are developing two interoperable toolkits for measurement and visualization. Autopilot, our distributed performance measurement and resource control system, is based on our experiences using the Pablo performance toolkit for application measurement [5,6] I O characterization [10], and WWW traffic analysis. Autopilot is complemented by Virtue, an immersive environment that accepts real time performance data from Autopilot and allows users to change software behavior and resource policies by directly manipulating representations of software structure and dynamics. In ....

Smirni, E., and Reed, D. A., "Workload Characterization of Input/Output Intensive Parallel Applications," Proceedings of the Ninth International Conference on Modeling Techniques and Tools for Computer Performance Evaluation, June 1997

....under NASA Contract NAG 1 613. achieving high performance for applications with large input output components. Moreover, most current parallel file systems were constructed as extensions of workstation file systems and optimized for large, sequential data transfers. Recent experimental studies [14, 2, 22, 23, 16] have shown that parallel applications have much more complex access patterns, with greater spatial and temporal variability, than first suspected. Although there is a large, complementary body of experimental data on disk behavior [19, 18] for sequential file systems, there is much less ....

....a small sample of the possibilities and attempting to extract more general patterns. Though the benchmarks and MESSKIT chemistry application we studied on the Intel Paragon XP S are but a few samples from a large space of possible input output patterns, earlier application characterization studies [2, 22, 23, 16, 14] suggest that our selections are representative of current practice. Although a wider range of experiments is desirable, the level of instrumentation and experiments we conducted required access to the operating system code and single user time to load experimental operating system kernels. This ....

Smirni, E., and Reed, D. A. Workload Characterization of Input/Output Intensive Parallel Applications. In Proceedings of the 9th International Conference on Modelling Techniques and Tools for Computer Performance Evaluation (June 1997).

....via global pointers from the sensor actuator manager. By changing the mapping, one can apply the rule set using different sensors, choose different policies, or even control different systems. 8 Parallel Input Output Example Our recent characterization studies of parallel input output patterns [3,18,24,25,23,22] have shown that parallel applications exhibit a wide variety of input output request patterns, with both very small and very large request sizes, sequential and non sequential access, and a variety of temporal variations. Small input output requests are best managed by aggregation, prefetching, ....

Evgenia Smirni and Daniel A. Reed. Workload Characterization of Input/Output Intensive Parallel Applications. In Proceedings of the 9th International Conference on Modelling Techniques and Tools for Computer Performance Evaluation, June 1997.

....that can estimate data access collisions. 4.3. Input Output For many high performance applications, the input output barrier rivals or exceeds that for computation, making design of scalable input output architectures and le systems a key aspect of any highperformance system. Our experience [7] has shown that software layering is a potential source of performance loss. Hence, we are developing an instrumented versions of multilevel input output libraries that enable multilevel analysis of I O requests (i.e. requests from the PACI ASCI Hierarchical Data Format (HDF 5) library, the ....

....Building on our input output characterizations, we are also developing queueing models of input output scalability and le striping. These models predict the performance of parallel disk systems as a function of request rates and access patterns, request sizes, and disk hardware parameters [7]. They can be used to estimate I O system scalability. 4.4. Networks and Computational Grids With increasing development of applications for heterogeneous, distributed computing grids, performance measurement and prediction must include application tuning for heterogeneous resources. The ....

E. Smirni and D. A. Reed. Workload Characterization of Input/Output Intensive Parallel Applications. In Proceedings of the 9th International Conference on Modelling Techniques and Tools for Computer Performance Evaluation, June 1997.

....parameters or resource policies during execution. Driven by the output of decision procedures, actuators have many of the same properties as sensors, including local computation and distributed activation. 3.4. 2 Adaptive Parallel File Systems As our extensive analysis of input output dynamics [6, 35, 32] has shown, the parallel input output patterns in emerging applications are both irregular and dynamic. Because the interactions between these applications and the file system software change during and across application executions [19] it is difficult or impossible to determine a globally ....

Smirni, E., and Reed, D. A. Workload Characterization of Input/Output Intensive Parallel Applications. In Proceedings of the 9th International Conference on Modelling Techniques and Tools for Computer Performance Evaluation (June 1997).

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