All of the test result graphs shown in this section include the performance of the platform during the user login, the steady-state phase once all the users have logged in, and the performance during the user logout phase. All of the graphs have time shown on the X-axis and the relevant metric displayed on the Y-axis.
The following table summarizes the test results for the various workloads and configurations that we tested:
Table 11. Workloads and configurations
Workload
Density per host
Average CPU
Average memory consumed (GB)
Average memory available (GB)
Network usage per host (Mbps)
IOPS per user
Task Worker
416
58.5%
480.4
543.2
2.01
4.2
Knowledge Worker
290
57.8%
742.3
285.4
8.21
31.5
Login VSI Task Worker Workload
The total user density achieved for this test was 1,250, with 416 user sessions per host. We deployed 14 multisession Azure Virtual Desktops on each AX-750 node to enable 1,250 Task Worker user sessions. The stack that we used was a 3-node Azure Stack HCI cluster, which hosted 42 multisession Azure Virtual Desktop VMs. There were 29.7 sessions per multisession VM, meaning that there were 3.7 user sessions per vCPU core.
Login VSI was configured to launch 1,250 user sessions in 60 minutes and to keep running the test for 30 minutes after the last user had been launched.
Three tests were performed with 1,250 user sessions. Only eight sessions failed on average during the test, which is an acceptable rate.
CPU usage
CPU usage steadily increased during the login phase and converged at the defined threshold during the steady state phase of the test. The compute nodes showed only slight discrepancies in usage during the login phase, which appears to be a normal distribution of usage. No anomalies were seen during testing. Note that due to data collector issues when running tests, the hosts were rebooted before testing. This reboot can lead to a lower CPU steady state average than during subsequent test runs.
The following graph shows the CPU utilization during testing. CPU usage with all VMs powered on was approximately 1 percent before the test started. During the steady state phase, an average CPU utilization of 58.5 percent was recorded.
Figure 28. CPU usage
Memory utilization
Out of 1,024 GB of available memory per node, the maximum consumed memory was 481.4 GB. The maximum memory pages per second was 156.8. Consumed memory remained below the physical capacity installed on each host. In addition to this, the memory paging observed in this test was normal. However, we noticed a spike in memory paging on two of the hosts (Host 1 and Host 3 as shown in the graph) during the steady-state phase. This did not last for more than 2 minutes. No anomalies were seen during testing. The following two figures show the consumed memory and memory pages/sec. The consumed memory graph is flat as all the memory was reserved and there was no anomaly.
Network bandwidth steadily increased during the login phase and converged during the steady-state phase of the test. One of the compute nodes showed slightly higher network traffic during the login, steady state and logout phases, which does not appear to be abnormal. No anomalies were seen during testing. Network usage reached a steady state average of 838.2 Mbps. The steady state peak was 1,271.7 Mbps and the maximum host peak was 1,388.5 Mbps.
Figure 31. Network traffic
IOPS
The total IOPS generated were as expected. The slight spike during the steady-state phases was normal and no anomalies were noticed during testing. The maximum disk IOPS was 17,744. The steady state peak was 3,223.7 and the steady state average was 1,732.2.
Figure 32. IOPS
Disk latency
Disk latency on all physical nodes remained at an acceptable level and well below the defined threshold (20 ms). No anomalies were seen during testing. The maximum read + read latency was 0.67 ms and the average steady state latency was 0.61 ms.
The VSI base of 584 that we achieved in our testing is an acceptable score that falls within Login VSI's rating of “Very Good." For more information about Login VSI baseline ratings and baseline calculations, see VSImax baseline scores.
As indicated by the blue line in the following figure, the system reached a VSImax average score of 1,037 when 1,250 sessions were loaded. This value is below the VSI threshold score of 1,584 set by the Login VSI tool. During testing, VSImax was never reached, which typically indicates a stable system and a better user experience.
Figure 36. User experience
Login VSI Knowledge Worker Workload
Additional disk space was required for the Knowledge Worker workload to have a balanced test setup. The virtual machines used differencing disks, hence the need for extra disk space. This space was achieved by adding three more SSDs, each of 900 GB, to the host machines. The total user density achieved for this test was 870, with 290 single-session Azure Virtual Desktops hosted across each AX-750 compute node. Login VSI was configured to launch 870 user sessions in 60 minutes and to keep running the test for 30 minutes after the last user had been launched.
Three tests were performed with 870 user sessions and none of the sessions failed during the test period.
CPU usage
CPU usage steadily increased during the login phase and converged well below the defined threshold during the steady-state phase of the test. The compute nodes showed only slight discrepancies in usage during the login phase, which appears to be a normal distribution of usage. No were anomalies seen during testing. Due to data collector issues when running tests, the hosts were rebooted before testing. A reboot can lead to a lower CPU steady-state average during subsequent test runs.
CPU usage at the start of the test was 2 percent and reached 57.81 percent during the steady-state phase.
Figure 37. CPU usage
Memory utilization
Out of 1,024 GB of available memory per node, the maximum consumed memory was 748.5 GB. The maximum memory pages per second was 22.6. Consumed memory steadily increased during login as expected and remained stable during the steady-state phase. Consumption started decreasing during logout. It remained below the physical capacity installed on each host. Memory paging during this test was normal and the spike in memory paging was not alarming.
Network bandwidth steadily increased during the login phase and converged during the steady-state phase of the test. The compute nodes showed only slight discrepancies in bandwidth during the login phase, which appears to be a normal distribution of network traffic. No anomalies were seen during testing. Network usage reached a steady state average of 2,380 Mbps. The steady state peak was 4,034 Mbps and the maximum host peak was 4,305 Mbps.
Figure 40. Network traffic
IOPS
The disk transfers/IOPS were as expected and no anomalies were found during the test. The maximum disk IOPS was 18,418. The steady state peak was 17,836 and the steady state average was 9,133.
Figure 41. Volume IOPS
Disk latency
Disk latency on all physical nodes remained at an acceptable level and below the defined threshold (20 ms). No anomalies were seen during testing. The maximum read + read latency was 3.5 ms and the average steady state latency was 1.02 ms.
The VSI base of 774 that we achieved in our testing is an acceptable score that falls within Login VSI's rating of "Very Good." For more information about Login VSI baseline ratings and baseline calculations, see VSImax baseline scores. The VSI index average shows that the system responded appropriately as additional load was added and the system behaved in a predictable manner. The response times were tightly grouped, meaning that there was a consistent user experience across VDI sessions.
As indicated by the blue line in the following figure, the system reached a VSImax average score of 1,028 when 870 sessions were loaded. This value is below the VSI threshold score of 1,774 set by the Login VSI tool. During testing, VSImax was never reached, which typically indicates a stable system and a better user experience. Figure 45. User experience
