Solve Section
This section contains the parameters and controls for actually calculating the USMN solution. They are summarized below:
-
Auto Solve. This button performs a “Best-Fit then Solve” (see below), trims outliers at 100% (see below), then performs a “Solve” (see below). This automated process is recommended for well-understood networks and in applications where the process has to be automated. It is not recommended when first optimizing an unfamiliar network of measurements. Anytime outliers are eliminated from the network, the causes and de- tails should be characterized and documented.
-
Do This Automatically. When the USMN dialog is displayed, an Auto Solve is automatically performed once.
-
Best-Fit Only. Performs a best-fit of the networked points. This is equivalent to performing a traditional best fit. Performing a best-fit prior to initiating the USMN solve drastically speeds up the calculation, because it provides for a much better initial starting point for the calculation.
-
Best-Fit then Solve. Performs a Best-Fit (see above) and Solve (see below) in one step.
-
Solve. Initiates a USMN calculation from the instrument’s cur- rent positions. R-Click to Enable/Disable CUDA devices processing. More on CUDA integration available in NVIDIA CUDA Devices.
-
Instrument Settings. Allows you to modify properties for all instruments at one time.
-
Trim Outliers. Removes points or observations exceeding a specified Ranking threshold from the USMN solution. Recall that ranking is an indicator of the quality of a point. Points above the ranking threshold can either be weighted to zero or removed from USMN entirely. Therefore, this is an effective means for preventing poor measurements from tainting the USMN solution.
-
Exclude Measurements. Any measurements that should not be included in the USMN calculation can be excluded with this button.
NVIDIA CUDA Devices
Using NVIDIA CUDA (Compute Unified Device Architecture) cores for USMN calculations provides an enhanced parallel processing option. It allows the use of the hundreds to thousands of cores contained within the GPU. On a decent NVidia graphics card (or alternately a Tesla module) this can significantly reduce the solve time for larger USMN networks. This requires a quality card and 64-bit SA.
There is an on/off option for this setting as a right-click function on the Solve button. This will activate CUDA devices if any are present and save the configuration to the registry so that it will persist between jobs. When CUDA devices are enabled/disabled, this transaction is logged so that the user can inspect the log for status.
Small USMN systems will not benefit from using CUDA devices. Each USMN solution sequence will take approximately 4 iterations to complete and the setup time to initialize and transfer data to the NVIDIA CUDA device simply adds more overhead time than is gained. So even if CUDA processing is enabled it will be ignored if an iteration takes less than 2 seconds.
Large USMN systems (30 instruments or more with 30 points or more per instrument) will benefit greatly from using NVIDIA CUDA devices.
A test showed the following results:
A network featuring 551 instruments with an average of 30 points per instrument and solving this system on a dual XEON (8 CPU cores) running at 2.39 GHz required 40 hours to solve (5 iterations at ~8 hours each). Solving this same system using NVIDIA CUDA devices (QUADRO P6000) reduced the solution time from 40 hours to 23 minutes.