Test results
Testing was conducted to examine how different hardware selections would impact editing workflow performance and user experience. Desktops were monitored as workflows were conducted under load.
Scripted testing was performed to simulate the steps an editor would take when performing the defined workflows. To provide meaningful results, all the system hardware and configuration (other than the desktop instances being tested) was held constant.
Upon test completion, results were assembled and analyzed to compare desktop utilization and end-user efficiency with different hardware configurations.
Impact of GPU configuration on desktop editing workflows
The following client configurations were used to compare the impact of a GPU on performance and user experience of the editing workflows on ArcGIS Pro:
- An Amazon EC2 R5XL instance (no GPU)
- An Amazon EC2 G4DNXL instance (GPU-enabled)
There are two sets of summarized results for each instance configuration (without GPU and with GPU) under each workflow.
Create a service
In this workflow, a new gas service endpoint was added to the network.
- Without GPU
- ArcGIS Pro 3.1 - Amazon EC2 R5XL instance (2 CPU / 4vCPU, 32 GB RAM)
- Workflow duration: 9.7 minutes
- Average CPU utilization: 48%
- Average memory utilization: 8 GB
- With GPU
- ArcGIS Pro 3.1 - Amazon EC2 G4DNXL instance (2 CPU / 4vCPU, 16 GB RAM, GPU - 16GB)
- Workflow duration: 8.5 minutes - reduced by 1.2 minutes (12%)
- Average CPU utilization: 38% - reduced by 21%
- Average memory utilization: 6.7 GB - reduced by 16%
Remove a service
In this workflow, a new gas service endpoint was removed from the network.
- Without GPU
- ArcGIS Pro 3.1 - Amazon EC2 R5XL instance (2 CPU / 4vCPU, 32 GB RAM)
- Workflow duration: 11.7 minutes
- Average CPU utilization: 58%
- Average memory utilization: 8.1 GB
- With GPU
- ArcGIS Pro 3.1 - Amazon EC2 G4DNXL instance (2 CPU / 4vCPU, 16 GB RAM, GPU - 16GB)
- Workflow duration: 9.0 minutes - reduced by 2.7 minutes (23%)
- Average CPU utilization: 45% - reduced by 22%
- Average memory utilization: 6.8 GB - reduced by 16%
Extend a main
In this workflow, a distribution pipe was added to the network.
- Without GPU
- ArcGIS Pro 3.1 - Amazon EC2 R5XL instance (2 CPU / 4vCPU, 32 GB RAM)
- Workflow duration: 10.0 minutes
- Average CPU utilization: 46%
- Average memory utilization: 8.1 GB
- With GPU
- ArcGIS Pro 3.1 - Amazon EC2 G4DNXL instance (2 CPU / 4vCPU, 16 GB RAM, GPU - 16GB)
- Workflow duration: 8.5 minutes – reduced by 1.5 minutes (15%)
- Average CPU utilization: 39% - reduced by 15%
- Average memory utilization: 6.8 GB - reduced by 16%
Replace a main
In this workflow, terminal connections were modified for a gas pipe.
- Without GPU
- ArcGIS Pro 3.1 - Amazon EC2 R5XL instance (2 CPU / 4vCPU, 32 GB RAM)
- Workflow duration: 16.0 minutes
- Average CPU utilization: 50%
- Average memory utilization: 8.4 GB
- With GPU
- ArcGIS Pro 3.1 - Amazon EC2 G4DNXL instance (2 CPU / 4vCPU, 16 GB RAM, GPU - 16GB)
- Workflow duration: 12.8 minutes - reduced by 3.2 minutes (20%)
- Average CPU utilization: 28% - reduced by 44%
- Average memory utilization: 7.1 GB - reduced by 15%
GPU workflow step times
While the system was under load, conducted workflow times across key workflow steps were captured. This represents the average time it took to complete a given step for both the instances with and without a GPU. Most steps are notably faster with a GPU enabled machine.
Beyond these key steps, the results across all workflows show a GPU-enabled instance is 20% faster and it provides a better user experience, improving the return on investment.
Conclusions for impact of GPU configuration
The R5XL instance (no GPU) experienced more events and wider peaks at 100% CPU utilization. In the GPU-enabled instance (G4DNXL), the GPU handled some of the processing, offloading work from the CPU. The workflow duration was shorter because the user was not waiting for the CPU. Additionally, the tests revealed a reduction in memory utilization with the G4DNXL instance as compared to the R5XL instance. This could be because the operating system needed to use additional memory as part of the GPU emulation processing.
The graph above shows the GPU (red line) handling some of the load as compared to CPU usage (orange area). The GPU was busy and sometimes exceeded the CPU usage, presumably during map rendering. This reduced the load on the CPU, provided a better user experience, and improved workflow times, as it was 19% faster across all workflows performed in this test.
Impact of CPU configuration on desktop editing workflows
The following client configurations were used to compare the impact of increasing desktops from 2 CPU/4 vCPU to 4 CPU/8 vCPU on performance and user experience of the editing workflows on ArcGIS Pro 2.9.5.
- An Amazon EC2 G4DN.XL instance (2 CPU/4 vCPU)
- An Amazon EC2 G4DN.2XL instance (4 CPU/8 vCPU)
Create a service
In this workflow, a customer gas service endpoint was added to the network.
- 4 vCPU
- ArcGIS Pro 2.9.5 – Amazon EC2 G4DN.XL instance (4 vCPU, 16 GB RAM, GPU-16GB)
- Average workflow duration: 8.2 minutes
- Average CPU utilization: 41%
- Average memory utilization: 6.7 GB
- 8 vCPU
- ArcGIS Pro 2.9.5 – Amazon EC2 G4DN.2XL instance (8 vCPU, 16 GB RAM, GPU-16GB)
- Average workflow duration: 7.8 minutes – reduced by 0.4 minutes (4%)
- Average CPU utilization: 16% - reduced by 61%
- Average memory utilization: 6.6 GB – reduced by 1.5%
Remove a service
In this workflow, a customer gas service pipe was removed from the network.
- 4 vCPU
- ArcGIS Pro 2.9.5 – Amazon EC2 G4DNXL instance (4 vCPU, 16 GB RAM, GPU-16GB)
- Average workflow duration: 8.7 minutes
- Average CPU utilization: 48.3%
- Average memory utilization: 6.7 GB
- 8 vCPU
- ArcGIS Pro 2.9.5 – Amazon EC2 G4DN.2XL instance (8 vCPU, 16 GB RAM, GPU-16GB)
- Average workflow duration: 7.9 minutes – reduced by 0.8 minutes (9%)
- Average CPU utilization: 18.6% - reduced by 60%
- Average memory utilization: 6.6 GB – reduced by 1.5%
CPU workflow step times
While the system was under load, conducted workflow times across key workflow steps were captured. They represent the average time it took to complete a given step for both instance sizes.
Conclusions for CPU configuration
Beyond the key steps, we looked at the total time for all steps in the four workflows tested. We observed that when increasing the instance size from 2CPU/4vCPU to 4CPU/8vCPU, the total time was 10% faster. One explanation for that result is the CPU usage shown in the chart below. Doubling the CPU allows ArcGIS Pro to further parallelize processing and improve overall processing efficiency, which reduced the average usage by an average of 63% across all workflows.
