Our testing consisted of 45 test cases based on the three graphics workloads (shown in Table 1) and the five network categories (shown in Table 7). We carried out three tests for each workload and network category combination (each test had 15 combinations) and averaged the results.
NVIDIA nVector Lite performance attributes
We ran the NVIDIA nVector Lite tool to assess the end-user experience. The tool measured the following remote user experience attributes:
- End-user latency: This metric defines the level of response of a remote desktop or application. It measures the duration of any lag that an end-user experiences when interacting with a remote desktop or application. The metric is based on the graphics driver on the user’s endpoint device and is measured along the whole graphics pipeline from the VDI desktop to the remote screen.
- Frame rate: This metric is a common measure of user experience and defines how smooth the experience is. It measures the rate at which frames per second (fps) are delivered on the screen of the endpoint device.
- Image quality: This metric defines the impact of remoting on image quality. It uses the Structural Similarity Index (SSIM) to compare an image that is rendered on the target desktop/workstation VM with the image that is displayed on the endpoint device. The average SSIM index of all pairs of images is computed for a single point in time for the remote VDI session. The index score is calculated once during the workload so a single value score is given for each workload.
WANem parameters
We configured the WANem emulator with the parameters for each network category as shown in Table 7 and configured the VDI desktop/workstation profiles as shown in Table 6. We applied the required network routing paths (routing using the WAN emulated network) on the appropriate infrastructure devices (such as the VDI desktop, the Horizon connection server, and the endpoint device) as shown in Figure 4.
Workloads
We selected representative user workloads for each graphics-intensive application as shown in Table 1. For all test instances, the workload we selected was chosen to represent the requirements of a medium user.