Toward an ideal data placement scheme for high performance object-oriented database systems

Abstract

Performance in parallel database systems very much depends on how data is placed on a network of processing elements. In designing a data placement method, particularly for high performance parallel object-oriented databases, two major factors that most of the time contradict each other, particularly minimizing communication cost and maintaining load balance, must be taken in account. Our experiments' results show that it is not feasible to trade I/O balancing with low communication costs, although these two factors are critical. Therefore, an ideal data placement method must feature a reasonable I/O balancing combined with low communication whenever possible, not the other way round.