The test was run in three parts.
For the first phase, we performed a single-stream minimal configuration test.
- For this test, we started with the vendor recommended setup of cards.
- We started the data collection and analysis with logging enabled and monitored the system performance using the Imagus data analysis system to measure the processed fps per stream and the frame processing time per stream.
- The GPU Profiler was used to measure the system and GPU resource utilization.
- The first phase of the testing was performed using local storage.
The second phase of was designed to evaluate the performance of the system using multiple VMs and the vGPU features of the NVIDIA GRID compute model.
- Testing involved setting up a single VM with two camera streams and allocated 25% of the NVIDIA T4 resources using vCenter and the NVIDIA vGPU driver.
- Identically configured VMs were added one-by-one to evaluate the effect of adding VMs as a way to increase capacity and provide for the possibility of a multi-tenant environment.
The third phase of the testing involved testing the failover time for a system with VMs configured for HA failover.
- Two host computers were prepared with identical resources. Each system had 3 NVIDIA T4 GPUs configure to be used as vGPUs using NVIDIA GRID.
- Each host contained the same compute cores and memory configuration
- Both hosts used shared storage for data store and heartbeat store.
- Both hosts were on the same network vLAN.
- A migration test was run first to establish that communication and resources were matched for the failover. The time of migration for a running VM was determined.
- The HA failover was configured in vCenter for the matched hosts.
- Power was removed from a running VM on one host and the time to failover (boot the VM and establish a running application) to the partner host was determined.