Models with a large number of elements produce very large matrices. The biggest single issue for the solution time as models become very large is how efficiently the software and operating system are able to use memory required for these matrices. A lot of memory swapping, especially writing to a hard drive becomes the single largest rate limiting step for large models. From the menu Solution>2D Triangular Elements>Problem Size you can see in the Message Area how much memory is required for a model in its current state.
For the example shown to the right, if Windows does not provide the memory required, the INTEGRATED program will require up to 7.983 GB for one scratch file and up to 2.153 GB for another. You should configure your software to use the fastest hard drive on the system for these files. Sometimes users have such large files that they assign them to different drives. However, the ideal solution is to keep all memory usage in RAM, or at least as much as possible.
For benchmarking purposes, the model referenced above was solved on two systems with varying amounts of installed RAM. The results are shown below for version 7.0 (the current release version as of March 2009).
Note some critical points from this plot:
- The trend is that more installed memory leads to more efficient solving, up to the amount of RAM the INTEGRATED program indicates is required for solving. This confirms that the rate limiting factor was the memory. After sufficient memory is installed, then the next issues will be CPU speed and multiprocessing (requires version 8.0). A user wishing to buy a system to solve large models more efficiently should note this amount of RAM for the largest models, and plan the purchase accordingly.
- The 32 and 64 bit systems vary in the amount of memory it is possible to install. The 32 bit system is much more limited.
- The two studies are only able to overlap on the 4 GB of RAM point. Here the 64 bit system is significantly faster, in spite of being a 20% slower CPU (see configuration information below). Accounting for the CPU speed, the fact that operations use 64 instead of 32 bits at a time is leading to an automatic speed improvement by a factor of approximately 2.
If an INTEGRATED user has a similar 10 GB problem to solve, and is presently using a 32 bit system with 2 GB of RAM, they can expect to solve approximately 7 times faster with a similar speed 64 bit system with 16 GB of RAM. In fact, the larger the memory required by a model, the more benefit there will be to using a 64 bit system. Customers with very large models have reported improvements by a factor of 10 to 30.
NOTE: INTEGRATED programs come in 32 and 64 bit installations. To obtain the benefits cited above you must not only have a 64 bit version of Windows, you must install a 64 bit build of the software.
APPENDIX: System Details
The above benchmark was done in version 7.0, prior to multiprocessing capability. Hence, the 2 CPUs on Test Series 2 is of no consequence to the results.