The cost model that corresponds to the temporal join processing model of section 8.3 is created in a similar approach as the one taken by Hua et al. when they model the performance of a parallel hash join in [Hua et al., 1991]. Hua et al. have shown that with their approach they were able to derive many interesting characteristics and simulate many developments that arose in real-world applications. We have already seen one example in figure 8.6. Another is the convergence towards hybrid parallel architectures - a development that was still neglected in 1992 by DeWitt and Gray in [DeWitt and Gray, 1992] but which has become reality nowadays (see discussion in [Norman et al., 1996]). We can therefore expect equally viable results when following their approach.
Thus we assume the hybrid architecture of section 8.2.3 and expect the temporal join to be processed as described in section 8.3. The costs are measured in seconds. The total response time of the temporal join depends on the times and spent in stages 1 and 2. In reality there might be an overlap between these two stages; thus
In our model, however, we assume that there is no overlap (e.g. enforced through a barrier type synchronisation). Thus we use the upper bound The stages (a), (b) etc. within stages 1 and 2 are treated accordingly, i.e.Furthermore we assume that the overlap between the I/O, communication, CPU and memory access phases within each stage is perfect. In reality, this can almost be achieved by separate I/O and communication processors. This means that we have to analyse the costs for I/O, communication, CPU and memory accesses for each substage of stages 1 and 2 and assume the maximum of these partial costs to be relevant for that substage. For stage 1 (a), for example, this means that
Section 8.4.2 describes how is calculated; section 8.4.3 does the same for . This analysis assumes certain parameters like I/O bandwidth, processor speed, amount of memory etc. These are introduced within those sections. As a convention we will use