Yasuhiro Hirano

Load balancing algorithms for parallel database processing on shared memory multiprocessors

This paper describes new load balancing algorithms for parallel database processing on shared memory multiprocessors. The goal of load balancing is to reduce overhead as well as load imbalance, but there is a tradeoff between them in ordinary algorithms. Unfortunately, optimum performance can hardly be obtained using ordinary algorithms because their performances depend on several factors such as database size, the number of processors and data distribution. The proposed algorithms solve these problems by varying the number of tasks allocated at a time ( which was fixed in ordinary algorithms ) according to the number of remaining tasks and the maximum and minimum processing times of a task. Performance evaluations show that the proposed algorithms achieve fair load balancing with lower overhead independent of the above factors.<<ETX>>

Extendible hashing for concurrent insertions and retrievals

Proposes an improved extendible hashing and bucket multi-versioning method, achieving a higher concurrency. In our improved extendible hashing, the global depth and directory entries are asynchronously modified to reduce lock conflicts on the directory. Furthermore, bucket multi-versioning enables read-only access to a bucket which is being split. Simulation studies show that these two methods provide speedup in proportion to the number of processors and enable concurrent insertions and retrievals to be performed without either one affecting the other.

Improved extendible hashing with high concurrency

We propose an improved extendible hashing scheme and a bucket multi-versioning scheme that achieve a higher concurrency. In our improved extendible hashing, the global depth and directory entries are asynchronously modified to reduce lock conflicts on the directory. Furthermore, bucket multi-versioning enables read-only access to a bucket which is being split. Simulation studies show that these two methods provide an increase in speed in proportion to the number of processors and that they enable concurrent insertions and retrievals to be performed without either affecting the other.

Load-sharing algorithms for parallel database processing on shared-everything multiprocessors

This paper describes new load-sharing algorithms for parallel database processing. There is a trade-off between overhead and load unbalance in ordinary algorithms. The proposed algorithms solve the tradeoff by varying the number of tasks allocated at one time, which is fixed in ordinary algorithms. Performance evaluations show that the proposed algorithms achieve fair load sharing with low overhead, independent of database size, the number of processors and data distribution.