Design of G-CODE algorithms and components

本研究旨在为网格计算开发环境（G-CODE）系统开发一个优化的网络感知调度程序所需的算法和协议。调度程序可以使用发送端算法和历史经验来估计带宽。它将使用启发式技术，如遗传算法，在近实时的情况下获得接近最优的调度。90年代初，采用自相似过程和长距离依赖关系的交通模型。然而，2000年进行的研究表明，交通量遵循泊松分布。因此，我们提出研究一种基于泊松分布的自诱导发送端算法。元调度器允许提前预约和基于工作流的任务调度。我们建议调查遗传算法的使用直到最低级别，其中它的使用可能取决于应用程序的并行深度。如果资源用动态图表示，如果每个节点用元组表示，如果应用程序用双权图表示，优化过程将在考虑通信延迟、延迟、先验保留和客户端能够建立的信任之后，将应用程序进程的需求与异构节点的计算能力相匹配。工作的新奇之处：没有报告的调度程序具有发送端带宽算法，并且能够处理工作流类型的工作，事先预留或通信延迟。当需要映射上千个进程时，调度器算法变得太慢了。没有人试图在每个层面上持续评估信任因素。此外，我们的研究正试图开发算法和协议，以与开放网格计算标准的移动目标保持一致。预期意义：如果我们要朝着全球网格的愿景前进，这里提到的新方法值得探索，无限数量的无处不在的设备与之相连。

The research aims at developing algorithms and protocols required for an optimized network- aware scheduler for the Grid Computing Development Environment (G-CODE) system. The scheduler may use sender-side algorithm and historical experience to estimate bandwidth. It would use heuristic techniques like genetic algorithm to obtain a near-optimal schedule in near-real time. In early nineties, the traffic models using self-similar processes and long range dependencies were used. However the studies conducted in 2000 have shown that traffic follows Poisson distribution. We, therefore, propose to investigate a self-induced sender-side algorithm based on Poisson distribution. The Meta scheduler is to permit advance reservations and work-flow based task scheduling. We propose to investigate the use of genetic algorithm up to the lowest level, where its use may depend upon the depth of parallelism of the application. If the resources are represented by a dynamic graph, if each node is represented by a tuple and if applications are represented by graphs with dual weights, optimization process will match the requirements of the processes of an application with the compute-powers of  the heterogeneous nodes, after considering the  communication delays, latencies, prior reservations and the trust that the client is able to build. NOVELTIES OF THE WORK: There is no reported scheduler with sender-end bandwidth algorithm and with the ability to handle workflow type of jobs, prior reservations or communication delays. The scheduler algorithms become too slow, when even a thousand processes are required to be mapped. None of them tries to continuously evaluate trust factor at every level. Moreover our research is trying to develop algorithms and protocols, which are in alignment with the moving target of open Grid Computing standards. EXPECTED SIGNIFICANCE: The new approaches mentioned here are worth exploring if we are to progress towards the vision of a global grid, with unlimited number of ubiquitous devices connected to it.