A Hybrid Architecture

Regarding the characteristics of the preceding three architectural types it becomes obvious that there is no ideal, single consensus architecture for parallel database systems. Hua et al. proposed an architecture that combines the advantages of shared-nothing (scalability) with those of shared-memory (fast communication, easy load-balancing) [Hua et al., 1991].   symmetric multiprocessor  (SMP)  nodes, each of which comprising   processors, are connected in a shared-nothing manner. The architecture is shown in figure 8.7. Many recent commercial products adopted this or similar architectures, such as the one outlined in figure 8.8: basically it is the same architecture as in figure 8.7 but allows access to disks from more than one node. As well as flexibility, this provides a certain redundancy in case that a node fails.

We will use this basic architectural model for modeling the performance of temporal joins. Hua et al. have proved its suitability for this purpose when analysing parallel join performances on this architecture and accurately simulating and predicting many architectural effects. We have already seen one example in figure 8.6. But this architectural model provides further advantages: the parameters and can be used to set up any of the architectural types that have been discussed so far:

• , : In this case, we have one processing node that contains one processor, i.e. a single-processor machine as it can be found in mainframes or other database servers.
• , :   This is a SM-/SD-architecture: one SMP node comprising processors.
• , :   This is a SN-architecture with single-processor nodes. Alternatively, it can be considered as a NUMA-based SM-/SD-architecture in which the interconnect represents a bus over which processors access non-local memory modules.