LS-DYNA Communication Performance Studies
In recent years, MPP-DYNA, the message passing parallel version of LS-DYNA, has become more and more popular in car crash and metal stamping simulations due to its good scalability which may reduce the turn-around time significantly when more processors are used. However, so far, most users only use 16 or less processors for LS-DYNA simulation because of the limitation of the scalability on a larger number of processors. This paper analyzes the communication patterns, message sizes and costs of simulation of two models. It is concluded that the unbalanced work load among processes is the bottleneck for scalability. Our study shows that some special decomposition techniques including sliding interface decomposition and scaling on certain directions may produce more balanced work load and, therefore, improve scalability. It is our hope that this study provides some insight for the algorithm improvement which may lead to better MPP-DYNA scalability on a larger number of processors.
https://www.dynalook.com/conferences/international-conf-2004/12-1.pdf/view
https://www.dynalook.com/@@site-logo/DYNAlook-Logo480x80.png
LS-DYNA Communication Performance Studies
In recent years, MPP-DYNA, the message passing parallel version of LS-DYNA, has become more and more popular in car crash and metal stamping simulations due to its good scalability which may reduce the turn-around time significantly when more processors are used. However, so far, most users only use 16 or less processors for LS-DYNA simulation because of the limitation of the scalability on a larger number of processors. This paper analyzes the communication patterns, message sizes and costs of simulation of two models. It is concluded that the unbalanced work load among processes is the bottleneck for scalability. Our study shows that some special decomposition techniques including sliding interface decomposition and scaling on certain directions may produce more balanced work load and, therefore, improve scalability. It is our hope that this study provides some insight for the algorithm improvement which may lead to better MPP-DYNA scalability on a larger number of processors.
12-1.pdf
— 129.0 KB