A Study of the Efficiency of Spatial Indexing Methods Applied to Large Astronomical Databases

We report the results of a study to compare the performance of two common database indexing methods, HTM and HEALPix, on Solaris and Windows database servers installed with PostgreSQL, and a Windows Server installed with MS SQL Server. The indexing was applied to the 2MASS All-Sky Catalog and to the Hubble Source Catalog, which approximate the diversity of catalogs common in astronomy. On each server, the study compared indexing performance by submitting 1 million queries at each index level with random sky positions and random cone search radius, which was computed on a logarithmic scale between 1 arcsec and 1 degree, and measuring the time to complete the query and write the output. These simulated queries, intended to model realistic use patterns, were run in a uniform way on many combinations of indexing method and indexing depth. The query times in all simulations are strongly I/O-bound and are linear with number of records returned for large numbers of sources. There are, however, considerable differences between simulations, which reveal that hardware I/O throughput is a more important factor in managing the performance of a DBMS than the choice of indexing scheme. The choice of index itself is relatively unimportant: for comparable index levels, the performance is consistent within the scatter of the timings. At small index levels (large cells; e.g. level 4; cell size 3.7 deg), there is large scatter in the timings because of wide variations in the number of sources found in the cells. At larger index levels, performance improves and scatter decreases, but the improvement at level 8 (14 arcmin) and higher is masked to some extent in the timing scatter caused by the range of query sizes. At very high levels (20; 0.0004 arsec), the granularity of the cells becomes so high that a large number of extraneous empty cells begin to degrade performance.