Latest revision as of 17:29, 4 April 2017

This section documents the improvements that were made to the NERPM-AB model since the adoption of the model as part as the “Path Forward 2040” Long Range Transportation Plan. The improvements focused on improving its usability related to runtime, user’s catalog interface, and model usability

Runtime

The investigation of alternative methods to reduce system run times focused on the highway network assignments and global feedback iterations. After investigating several different options for reducing runtime the following improvements were made:

Feedback Loops

The transition of model outputs from one feedback loop to the next one and feedback loop termination criteria were adjusted in the new model. The model has been set up such that the weighted-average volume for each link is calculated at the end of each loop, based on the current and previous feedback loop flows. These averaged volumes are used to generate skims for the next iteration of the model system and are fed back to the demand models (DaySim).

In addition, rather than terminate based on a fixed number of iterations, the model has been set up to terminate the feedback loops if both of the following criteria are met:

1. Change in total AM and PM VMT of two successive feedback loops is less than five percent
2. The AM and PM volumes of more than 95 percent of freeway and arterial links change less than five percent in two successive loops.

The net effect of these changes are that the model runs one feedback loop less than the original model for the 2010 baseline scenario.

Highway Traffic Assignment

The model is able to run midday and nighttime highway assignments in parallel. The user can assign computer logical processors to time periods by filling the corresponding keys added to the user interface. Figure 7-1, below, shows an example of the Cube catalog (last page of user interface page) for a computer with 8 logical processors. The AM and the PM assignments use all available logical processors, whereas the midday and nighttime assignments use different logical processors and can therefore run in parallel.

The model currently runs AM and PM assignments in sequence; however, it will be able to run them in parallel with minor changes in the model. In this case, the model should know which computer logical processor is the first one used for AM assignment (also which one is the first one for PM period). The interface has these two keys now and, since they are not run in parallel, the value for them should be higher than number of computer logical processors. The default value, therefore, was set to 25, which is suitable for any machine with up to 24 logical processors.

Figure 7-1: Assigning computer logical processors to highway assignments by time period

Highway Traffic Assignment Parameters

The assignment convergence criterion was changed to the accepted standard practice of relative gap. In addition, the maximum number of iterations was increased from 50 to 150, allowing for greater accuracy in the assignment process. The model is setup to use a tighter relative gap in the highway assignment after the first loop. Following best practice recommendations for regional highway assignments, it is recommended that this relative gap be set to 0.0001. The user may increase or decrease the relative gap, depending on the accuracy level required for the analysis under consideration. Figure 7-2 shows that the relative gap used after the first feedback loop was added to the interface as a key controlled by the user.

Figure 7-2: Setting the relative gap used in highway assignment after the second feedback loop

Post-Process Analysis

The Select link analysis is set to run after the feedback loop. The user selects the option to run a selected link analysis on typically the third page of the user interface, as shown in Figure 7-3. The select link analysis can be run for any of the time periods, including the daily assignment, by selecting the specific option. In order to run the daily select link option, all the time periods need to be run for the select link.

If the user wants to save the congested path files, they can choose the option to save the path file for a particular period. The paths are saved in an extra highway assignment run at the end of the model based on the final congested network. It should be noted that Cube Cluster does not support the use of multiple logical processors for path saving; therefore, this step takes a bit longer than would a normal assignment that can be distributed across processors.

Figure 7-3: Keys used to define links for select link analysis and time periods for path saving

Table 7-1 lists all the fields in the loaded network files. The two columns on the right lists the fields associated with the links which the user selected in the "select link" option shown in Figure 7-3. The yellow rows on the table are used in the calculation listed on the last row of the table and calculates the total link volume.

Table 7-1 NERPAB1v2 Loaded highway network fields

Redundant Steps

Several steps in the trip distribution part of the model are no longer needed.

Model operations previously used to combine trip tables are no longer necessary because DaySim uses its own set of procedures for creating trip tables from trip lists.

The pre-assignment step is now run only once, during the first feedback loop iteration. This is because the outputs from pre-assignment are used to build the transit network, which is only run during the first feedback loop iteration.

Clean up Cube Catalog Interface

The validation base year key (Valyear) is not read by the model and was removed from the catalog interface. All other keys in the interface are needed.

Performance Improvements

Table 7-2 shows the run times based on a 2015 vintage workstation with 8 logical processors before the changes were made and the run times after the changes, with various settings. Comparing the rows shaded in orange, the changes made to the model decreased the run time by more than 9 hours, a 44 percent improvement, for the base year model (2010), using the stricter relative gap and without path saving and select link analysis. The run of the same model with select link analysis for AM and PM periods and saving path file for AM period ran for 16:27, using the stricter relative gap, which is reasonable.

Table 7-2: Run time of the original and improved (V2) NERPMAB1v2 models Using a workstation with 8 logical processors

New and Updated Scripts

Several new scripts were created and/or updated to support the listed changes. Table 7-3 shows the four scripts that are updated in the NERPMAB1v2 model, all corresponding to the highway assignments by time period. Table 7-4 presents the new scripts in the NERPMAB1v2 model with a brief description for each.

Table 7-3: Updated Scripts in the NERPMAB1v2 model

Table 7-4: New scripts in the NERPMAB1v2 model