CrossGrid Developer Manual Guide

Introduction

The PPC (Performance Prediction Component) is a tool to study the behavior of some selected MPI kernels in the Grid. It is based on a set of models that characterize the behavior of the kernels when they are executed in the Grid. These models were obtained from exhaustive executions on different Grid scenarios. From these measurements a methodology to obtain correlations between performance and monitoring information is applied to obtain easy to use functions to characterize the execution of the kernel. Even though the tool is limited to a set of specific kernels, it can be easily extended to others. Note that some of the kernels are general purpose. i.e. communications, that could be useful for any user.

The tool shows performance information based on predictions taken form the models. The way to show this information is interactive and easy to use. Also the tool can get monitoring information provided by JIMS to show the predictions under the present and real scenario, and then the user can interactively change these values to analyze their effects on the kernel.

Some of the main features of the tool are:

• PPC is an application dependent tool. It is specifically applied to selected kernels that were analyzed in terms of performance to obtain analytical models. Anyway, it can be also broadly used to study the performance of communication routines themselves.
• PPC is an interactive tool in the sense that the user can easily change the parameters that characterize the system or the problem, and then analyze their influence.
• PPC uses analytical models, and therefore the predictions are obtained fast. This feature is very important for the interactivity.
• PPC can be used to understand the behavior of the kernels under different Grid features, selected by the user. This is important in any system in which the features change dynamically, like Grids.
• PPC can be easily extended to any heterogeneous system and kernel.

One of the main contributions of PPC to the state of the art in Grid computing is the methodology to obtain the performance models, and to allow the user to understand the behavior of the kernels under hypothetical Grid features. Please refer to PPC User Manual for details.

Right-click here to download PDF version