hhhong@hyowon. cc.pusan. ac. k r
1, IntroductionSingle-scan queries, such as point query or region query, can bc efficiently processed by using spatial indices. Spatial joins, however, are required to do multi-scan of all of the spatial objects. This gives us bad impact because the total time o f executing a spatial join increases exponentially in proportion to the number of spatial objects [ 1,12,14]. Although scvcral spatial join algorithm recently have been dcvcloped, the spatial join still remains to be a hie-consiiming operation where tlie response time is far bcyond the acccptable spced of ifti intcractive iiser [3]. The molivation of this work is to analyze thc impact of parallel proccssing on porforining the spatial join.There are two categories of spatial Indices For processing spatial joins: single-asigntneiit and multiassignment[ 16,171. 'I'he singje-as.rignmei7f spatial index Iias the Single-Assignmunt, Multi-Join(SAMJ) property as one object is assigned into one and only one buckct, and 0-7696-0496-WOO $10.00 0 2000 IEEE the bitckct is required to join with multiple: buckets. A hucket contailis all the objects of which thc centroid of MBR is eticloscd within the bucker., For example, the object A is assigned to the bnckct 5 , the object B to the bucket 2 in iigure I . On the othcr hand, thc Iniiltirrssigniiient spitial index usually has the Multi-Assigntrient, Sirigle-Join(MASJ) propei-ty as onc object can be assigned to inultiple buckets, and any bucket always joins with only one bucket. For instance, whilc the object A is only assigned to the bucket 5 , the object B overlapping with four buckets is assigned to all the buckets I , 2, 4, and 5 in ti "' _e 1,In this papcr, we proposc two parallel spatial join algorithms opclated 011 two spatial indices(a singleassigiimcnt and n multi-assigamcnt grid filc) dcrived from tlic traditional fixed grid filc. 'J'lie first algorithm is the parallel spatial join processing by using a single assigntneiit grid filc. l'he second is the parallcl spatial join processing by using a iniiltiple assignment grid file. The parallcl spatial join algorithtns arc realized by using tlircc kinds of task allocation methods: static, dynamic and semi-dynamic. The setni-dynamic task allocation method newly proposed in this paper utilizcs tlie sizc as well as thc spatial localities of tasks for load balancing. The method makes iise of the merits of both static and dynamic load balancing.In ordcr to evaluate the perforrnaiice of the two parallel spatial join algorithins and thrcc task altocation methods, we conducted expcriments with a real spatial data set. Thc platform on which these tests are clonc is an IBM SP2 parallcl tnachinc having MIMD with sliarcd disk architecture and high bendwidth chantiel for rnessagc communication betwccn processors. Thc objective of this paper is to fiiid out which parallel spatial join algorithm and wliicti task allocation stratcgy based on grid files shows the best pcrforrrrance. Also, the effect of parallel processing ut' spatial join will be disc...