Técnicas de Coscheduling para la Computación en Clusters No Dedicados
Document information
| Author | Francesc Solsona Tehàs |
| School | Universidad Autónoma de Barcelona |
| Major | Informática |
| Year of publication | 2002 |
| Place | Barcelona |
| Document type | thesis |
| instructor | Dr. Porfidio Hernández Budé |
| Language | Spanish |
| Number of pages | 107 |
| Format | |
| Size | 350.81 KB |
- Técnicas de Coscheduling
- Computación en Clusters
- Procesamiento Distribuido
Summary
I. Introduction to Coscheduling Techniques
The document presents Coscheduling Techniques for Non-Dedicated Cluster Computing, focusing on the integration of general-purpose workstations interconnected via Local Area Networks (LANs). These workstations form a Network of Workstations (NOW), designed to execute applications with minimal computing requirements. The document emphasizes the distinction between local and distributed applications, highlighting that local applications are typically I/O bound, while distributed applications often require more complex systems. The significance of this research lies in addressing the challenges faced by clusters when executing both types of applications simultaneously. The author notes, 'The performance of distributed applications depends on the behavior of all their forming processes.' This statement underscores the need for effective resource management in cluster environments.
1.1. Overview of Cluster Systems
Cluster systems consist of multiple processing elements (PEs) that can either share a global memory or operate with distributed memory. The document discusses the classification of applications into parallel and distributed processing, emphasizing the communication methods used in each case. The evolution of Distributed Computing Environments (DCE), such as PVM and MPI, is also addressed, showcasing their role in simplifying the design and implementation of distributed applications. The increasing popularity of NOWs in distributed processing is noted, along with the challenges that arise when local and distributed applications are executed in parallel. The author states, 'There are important drawbacks to be considered if both distributed and local applications are executed in parallel in a Cluster system.' This highlights the critical need for effective scheduling techniques.
II. Challenges in Cluster Computing
The document identifies key challenges in Cluster Computing, particularly when local and distributed applications are executed concurrently. The performance of distributed applications can significantly decline if cluster nodes are overloaded with local jobs. Conversely, local tasks may experience increased response times due to heavy distributed workloads. The author articulates the need for a new research goal: 'how to coordinate local and distributed applications when they are executed in parallel in a Cluster system.' This statement reflects the urgency of developing effective coordination strategies to optimize performance across different application types. The document also discusses the implications of CPU-bound and I/O-bound tasks, emphasizing the importance of resource allocation in maintaining system efficiency.
2.1. Resource Management Strategies
Effective resource management is crucial for the successful execution of both local and distributed applications. The document proposes the construction of a Virtual Machine over a cluster system to facilitate the efficient execution of traditional workstation jobs alongside distributed applications. The need for time-sharing operating systems is highlighted, as they allow for the equitable distribution of CPU time among executing tasks. The author notes, 'To solve the problem, two major considerations must be addressed: how to share and schedule the workstation resources and how to manage the overall system for efficient execution.' This insight emphasizes the necessity of developing robust scheduling algorithms to enhance performance in cluster environments.
III. Coscheduling Techniques
The core of the document revolves around Coscheduling Techniques, which are essential for optimizing CPU resource sharing between local and distributed applications. The author explains that coscheduling aims to reduce communication waiting times by scheduling tasks concurrently. The document categorizes coscheduling into two main trends: explicit and implicit control. Explicit control involves specialized processes managing distributed tasks, while implicit control relies on local scheduling decisions based on real-time events. The author states, 'Coscheduling techniques follow two major trends: explicit and implicit control,' highlighting the flexibility of these approaches in adapting to varying workload conditions. The significance of these techniques lies in their ability to enhance the performance of distributed applications while minimizing the impact on local workloads.
3.1. Mechanisms and Features
The document presents two coscheduling mechanisms that embody the explicit and implicit control trends. These mechanisms are designed to provide usability, good performance in executing distributed applications, and low overhead. The author emphasizes that the design of these techniques was influenced by the optimization of key features, such as simultaneous execution and low impact on local workload performance. The introduction of an implicit-control coscheduling model is also discussed, which includes features like collecting on-time performance statistics. This model serves as a foundational scheme for enhancing the efficiency of cluster systems, demonstrating the practical applications of the research findings.
Document reference
- Coscheduling Techniques for Non-Dedicated Cluster Computing (Francesc Solsona Tehàs)
- Arquitectura de Computadores y Procesamiento Paralelo (Dr. Porfidio Hernández Budé)
- PVM (Parallel Virtual Machine)
- MPI (Message Passing Interface)
- Unidad de Arquitectura y Sistemas Operativos (Emilio Luque)