In the first four tests, we deployed and ran our HammerDB test harness against sqlmi-01 through sqlmi-04, independently. As expected, we observed near linear scale of TPMs as the number of users increased. Also, the greatest TPMs were driven by the maximum 25 users when accessing the SQL MI with the highest requested compute resources. Memory utilization was never under pressure during the testing.
For these tests, we set the vCPU count requested to half the vCPU limit. In some cases, we observed increased performance with this configuration versus setting vCPU requested equal to limit. The TPM numbers were still within an acceptable range considering the resources that were allocated and the running workloads.
With the combined workload profile, we saw the greatest efficiency using a SQL MI with 2 requested vCPUs and a 4-vCPU limit. We considered it a best practice to maximize work with a reasonable core count even using a modern consumption-based pricing model.
2 vCPU, 4 Gi
4 vCPU, 8 Gi