Reclustering of high energy physics data

The coming high energy physics experiments will store Petabytes of data into object databases. Analysis jobs will frequently traverse collections containing millions of stored objects. Clustering is one of the most effective means to enhance the performance of these applications. The paper presents a reclustering algorithm for independent objects contained in multiple possibly overlapping collections on secondary storage. The algorithm decomposes the stored objects into a number of independent chunks and then maps these chunks to a traveling salesman problem. Under a set of realistic assumptions, the number of disk seeks is reduced almost to the theoretical minimum. Experimental results obtained from a prototype are included