the number of Aggregators and the number of levels in the aggregation tree.
The model presented in Section 4 was very useful in helping us
understand these scale up results.
5.4 Multi-Client Experiments
The final experiment tests query latency and throughput as the
number of clients increases from 1 to 32. Each client issues 200
consecutive queries with no time between them. For this experiment, we use 160 machines with 1 GB of data at each Leaf. We
also use the 3 hour, 6 hour, and 29 hour variants of both the scan
and time series queries. The 3 hour variant only needs to scan about
10% of the data at each leaf.
Figure 12 shows the throughput of each query as we vary the
number of clients. The first thing to notice is that for each query, the
throughput rises as the number of clients increases, until the CPUs
at the leaves are saturated. After that, the throughput flattens out.
For all queries, the throughput is flat after 8 clients. The next no