Grid Data Management: Simulations of Lcg 2008

Simulations have been performed with the grid simulator OptorSim using the expected analysis patterns from the LHC experiments and a realistic model of the LCG at LHC startup, with thousands of user analysis jobs running at over a hundred grid sites. It is shown, first, that dynamic data replication plays a significant role in the overall analysis throughput in terms of optimising job throughput and reducing network usage; second, that simple file deletion algorithms such as LRU and LFU algorithms are as effective as economic models; third, that site policies which allow all experiments to share resources in a global Grid is more effective in terms of data access time and network usage; and lastly, that dynamic data management applied to user data access patterns where particular files are accessed more often (characterised by a Zipf power law function) lead to much improved performance compared to sequential access.