Parallel database
systems are implemented on one of the alternative parallel architectures: sharedmemory,
shared-disk, shared nothing, hierarchical, or NUMA (Valduriez & Ozsu,
1999), which have implications on parallel query processing algorithms, data partitioning,
and placement. In practice, parallel environments involve several extra
overheads related to data and control exchanges between processing units and also
concerning storage, so that all components of the system need to be designed to
avoid bottlenecks that would compromise the whole processing efficiency. Some
parts of the system have to account for the aggregate flow into/from all units. For
instance, in shared-disk systems the storage devices and interconnections should be
sufficiently fast to handle the aggregate of all accesses without becoming a signifi-
cant bottleneck. To handle these requirements, a significant initial and continuous
investment is necessary in specialized, fast, and fully-dedicated hardware. An attractive
alternative is to use a number of low-cost computer nodes in a shared-nothing
environment, possibly in a nondedicated local network. The only requirement
is that each node has some database engine and connectivity, while a middle layer
provides parallel processing. This system must take into consideration partitioning
and processing, as the computer nodes and interconnects are not specially designed
to that end.
Pages:
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391