Grid computing is a work in progress

Computational grids are not new, but ongoing research suggests there is scope for improvement
Written by Alan Oxley
Computing, 08 Oct 2009

A grid is a computer network in which the resources are pooled. A job arriving on the grid needs to be allocated to a computer or, possibly, split up and allocated to several computers. The software responsible for making this decision must find a suitable computer. This activity is termed resource allocation. When jobs are queuing up awaiting the allocation of a computer, some queuing strategy needs to be employed.

This activity is termed job scheduling, for which an appropriate scheduling algorithm must be employed.

There is an almost infinite variety of algorithms that can be used. The final choice needs to reflect both the size, shape and nature of the grid’s architecture, and the traffic it is expected to handle.

Numerous recordings of traffic, known as workload traces, are available for researchers to study. By using statistical methods, it is possible to identify patterns present in a workload trace, patterns such as short-range dependence, long-range dependence, and self-similarity. By gaining a more detailed understanding of traffic patterns, it should be possible to generate any number of synthetic workloads against which algorithms could be tested.

Let us return briefly to another problem that we mentioned, resource allocation. To understand this topic, it helps to think in terms of an economy grid, where each computer is offering its services for a price.

Each computer site informs the grid manager – a program – about the resources that it has, its job turnaround times, its reliability, and the prices it is going to charge. The complexity of the situation is compounded by the fact that any computer on the grid pools only some of its resources, the rest being used locally.

The grid manager could assume that each site has told the truth about its likely performance. This would obviously be an unwise thing to do and so the grid manager must look at past performance as an indicator to future performance.

Therefore, we can appreciate that there is plenty for researchers to do in striving to improve the efficiency of computational grids.

from: http://www.computing.co.uk/computing/comment/2250683/grid-computing-work-progress-4836792