In recent years, INTEGRATED has pushed ahead with 64-bit processing and then with strategically rewriting more of the time-critical parts of the 3D software to enable multithreading (use of more than one processor core at a time). At the same time, computer manufacturers have made large amounts of memory and multiple processor cores quite affordable. As a result, many users have found dramatic improvements in the solving speed for their most difficult problem. However, the exact speed improvement is dependent on many factors and it is not always clear which factors will be most significant for a given type of simulation.
INTEGRATED has previously presented benchmarks for 32 versus 64-bit computing and the effect of the amount of RAM. We are now conducting a much larger study comparing these and number of Processor Cores. In this article we present some preliminary results.
The graph below shows the effect of solving a single standard problem under a variety of configurations for a 64 bit Windows 7 system. The hardware has 32 GB of RAM and 2 quad core processors. These were selectively reduced under the Windows Advanced Boot options to study the trends with everything else fixed. COULOMB reported that it requires 6 GB of memory in order to solve the selected problem.
Notice that there is an almost monotonic trend of solving faster as the amount of RAM or the number of processor cores is increased. As earlier benchmarks already showed, there is an especially big improvement as the amount of RAM approaches the reported amount of memory COULOMB reports needing, but we can also see in these graphs that the speed improvement is more substantial when both the number of cores and the amount of RAM is increased than when either is increased alone:
- With a small amount of RAM, 8 processors only solves 30% faster than one processor. With lots of RAM 8 processors solves 4x faster than one processor
- With one processor increasing the RAM produces a speed increase of about 30%. With 8 processors increasing the RAM produces a speed increase of about 4x.
NOTE: the terms “small” or “large” are used deliberately here to emphasize the principle. What is small or large depends on what amount of memory the INTEGRATED program indicates will be required for the model. Hence, it is impossible to properly gauge the best system without inquiring memory requirements for some sample models.
Qualifiers and Concluding Comments:
This study has been constructed to determine what should be good guiding principles for making hardware choices for use with INTEGRATED 3D software. The exact results will be dependent on the actual program and models, but the principles should be common. The main benefits of the type shown above will be found with the BEM solver in 3D programs. However, the same advancements will continue to be further expanded to other parts of the software in coming versions.