What troubleshooting options are available for a FEMFAT parallel analysis?

In many typical applications (e.g. multiaxial damage analysis of a body-in-white with several hundred channels and measured load-time histories), it is urgently recommended that the parallelization be used if the duration of the fatigue analysis plays a decisive role. 

FEMFAT employs a “model-based” parallelization for this purpose: starting with the imported geometry (and in ChannelMAX, also including consideration of the load) the FEMFAT main process (so-called “master”) creates analysis groups and separate job files. These are processed by separate instances of FEMFAT (so-called “slaves”) that are started additionally. 

The file parallconfig.txt (can be found in the FEMFAT installation folder under “Resources”) contains important parameters for the control of the FEMFAT parallel analysis. The two parameters of particular importance for troubleshooting are:  

  • slaveStartTimeout: max. waiting time (in sec.) of the master process until all slave processes have been initialized. Even if only one slave has not finished initialization by this time, the analysis is aborted.
  • slaveFinishTimeout: max. waiting time (in sec.) of the master process until the end of the analysis of the slowest slave.

Recommended procedure for problems:

a) Our basic recommendation is to check the msg, pro and/or fps files of the individual slave processes (in the subfolders of the FF parallel folder) as well as of the FEMFAT main process (typically in the working directory) for errors or warnings.

b) If these files are not available, or the typical directory structure for a FEMFAT parallel analysis does not exist, please check whether sufficient memory and licenses were available for the selected quantity of CPUs at the time of start-up.

  • The required hard disk space can be estimated by: Memory for non-parallelized analysis x number of CPUs 
  • The main memory requirement is very heavily dependent on the size of the model. The following rule of thumb has proven useful in practice: 3 GB RAM per 1 million nodes and CPU. 

This already includes extra memory for the system and any possible calls of the FE adapter.

c) No analysis progress although FF parallel folders are present and slave processes have been started. Please check whether the SlaveStartTimeout has been chosen so that it is sufficiently large (timeouts can result especially with large models and/or very many large scratch files).

d) Partial areas with no results: Check the fps, msg and pro files of the slaves for errors and check the filter settings and the outputs on the console.

e) Performance problems: Please bear in mind here that in FEMFAT, only the actual analysis is parallelized. Currently, neither the reading of the structure data nor the scratching of the stress data can be executed in parallel. 

Also bear in mind that time must also always be allowed for the splitting of the structure and the merging of the partial results. This time can be the dominant factor in the case of smaller models or short overall analysis times!

Tip: For optimal performance with Abaqus odb files we recommend either the use of the odb version that is consistent with the respective FEMFAT FE adapter version or odb files that have already been updated previously. Otherwise, every slave process carries out an update of the odb database.