To test the hyper-converged system, we used an incremental load test that was designed to distribute load across the four S2D nodes. For example, one virtualized database ran on each node in test 4, and two databases ran on each node in test 8.
The following figure shows the incremental test plan with the distributed workload.
Figure 2. Incremental test plan with distributed workload
We used Benchmark Factory and Dell EMC Live Optics to capture performance and utilization data. For more information, see Appendix B: Test approach and metrics collection.
For this paper we review transactions per second (TPS) because it is a commonly used database-performance metric. The hyper-converged platform showed near-linear scalability in the incremental load test scenario, as shown in the following figure.
Figure 3. Total TPS for each incremental load test
When one virtualized database was running on the platform, it achieved a peak of 1,823 TPS, and when all eight virtualized databases were running, the total TPS was 13,351 and the average TPS was 1,669. The difference is a minor drop of 154 IOPS for each of the eight virtualized databases when compared to the single virtualized database.
After we obtained the incremental load findings through eight virtualized databases on the hyper-converged platform, we modeled four additional databases. In this way, we could achieve a consolidation ratio of 3 to 1, meaning that we had three virtualized databases running on one physical S2D node.
To model the four additional databases, we had to account for a slight degradation in performance of the hyper-converged platform. If characterized correctly, the TPS slope shows how all-flash technology can sustain increased workloads with a minor drop in performance. The goal was to err on the slightly conservative side to provide room for error, meaning that the hyper-converged platform could perform better than our model. We used a 5 percent loss in TPS to model each additional virtualized database. The following figure shows the resulting chart.
Figure 4. Model of TPS with four additional databases
The orange data points indicate the four modeled databases. Overall, the slope of the four modeled databases is very close to the slope of the findings for the eight virtualized databases. Using this model, the estimated TPS for 12 databases running on the hyper-converged infrastructure is 19,368.