The next test case used 16 vCPU VMs in a scale-out test scenario. Each processor in the physical servers that were used for the SDDC had 18 cores. The 16 vCPU test case dealt with VMs that were just two cores smaller than the NUMA node size of the physical server.
In this round of tests, the scaling remains strong up through 12 VMs, but it begins to fall off with 13 (Figure 6). While the previous test with 8 vCPU VMs was limited by memory, this test was limited by CPU.
Figure 6. Scale-out performance with 16 vCPU Oracle/RHEL VMs running DVD Store 3
Each host had 36 cores, with the four-node SDDC having a total of 144 cores. With twelve 16-vCPU VMs, there were 192 vCPUs allocated, which means that the cloud had 48 more vCPUs allocated than it had physical cores.
Regardless, performance still scaled well because of hyperthreading, which gave the system a total of 288 threads. At 13 VMs, the total vCPUs reach 212 and performance degraded a bit, but all VMs continue to achieve significant throughput under this load. But at 14 VMs, the tests did not complete successfully. The load driver VM and all of the database VMs reach a total of 244 vCPUs assigned, and not all database VMs were able to get enough resources to complete queries in a timely manner.