Hironori Emaru

DGP: Dynamic grouping of particles for parallel molecular dynamics simulations

Molecular dynamics simulations (MD) have become very important in physics and chemistry. In this paper, we propose dynamic grouping of particles (DGP) for effective parallel processing of short-range MD. DGP is based on a cyclic process of grouping the particles that are within a registration distance of each other. Such a particle group defines a parallel task that can be processed independently during the registration step. We have implemented our method on SR2201, a loosely coupled parallel processor. Experimental results show that DGP is twice as fast as a conventional method based on cell partitioning.

Performance Management for the Dynamic Storage Tiering by Virtual Volume Clustering

In recent years, dynamic storage tiering has been proposed as one of the storage virtualization technologies. In this storage system, a volume is divided into sub-volumes, which can be migrated among different performance storage media according to the frequency of I/O operations, etc. This technology is expected to reduce the total storage media cost. But there is no method to assure a certain service level for the storage system. We discuss five requirements: (1) performance management of arbitrary volume granularity, (2) system scalability, (3) adaptability to unknown environments, (4) general versatility, (5) reducing a total storage media cost. To satisfy these five requirements, we propose a virtual volume clustering method and implement it in a simulation environment. By analyzing the simulation results, we summarize the three considerations for further improvement: (1) wide variation in target response time, (2) maintenance of service level, (3) migration of a large capacity volume. Based on these three considerations, we improve our proposal method. We show that our improved method is effective by an evaluation in a simulation environment.