Each
of these issues requires specific solutions. For instance, network failures can be accommodated
using backup connections. We concentrate on handling the possible
unavailability of computing nodes, guaranteeing efficient availability, and promoting
manageability. The objective is that the system be always-on and always ef-
ficient even when nodes are unavailable or entire parts of it are taken off-line for
maintenance and management functions, such as loading with new data or other
DBA functionality. Efficient node availability can be achieved via the use of replicas.
A replica is a ???standby??? copy of some data that can be activated at any moment
in case of unavailability or failure of the node holding the ???original,??? so that
processing resumes as usual. If processing with unavailable nodes is implemented
efficiently, unavailability becomes less onerous to the whole system and it becomes
feasible to have nodes unavailable or to stop a set of nodes for data loading, maintenance,
upgrading, or other management activities without any major repercussions
to processing. Replica placement has been studied in different contexts, from
RAID disks (Patterson, Gibson, & Katz, 1998) to the context of generic parallel
2 0 Furtado
Copyright ?© 2007, Idea Group Inc. Copying or distributing in print or electronic forms without written permission of
Idea Group Inc.
Pages:
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402