Maximizing Memory Performance for ANSYS Simulations
Memory, or RAM, is important when you’re configuring computers for high-performance computing (HPC). Computer performance is affected not only by the amount of RAM but also by the amount of bandwidth, or transfer speed, between the processor and the memory. The bandwidth is affected by the speed of the memory stick, but the way memory sticks are configured on the motherboard is even more important.
CPUs can access information on multiple memory sticks in parallel through use of memory channels. Modern CPUs can access between 1 and 8 memory channels simultaneously (model specific). Not only does this capability affect the machines performance, but we have found that many computers are not configured to make use of all their memory channels, which hampers performance.
One of the main reasons people witness disappointing performance gains when adding more cores to their simulations is not due to inefficiency of the solver code, but instead due to a combination of decreasing core frequency (turbo speed) and subdivision of out-of-core resources, especially memory bandwidth.
In this paper, SimuTech HPC Expert Alex Pickard shares our recent test on how memory channels can impact your simulation performance.
- CPU data starvation, and how to overcome it
- Our test of memory bandwidth via RAM configuration changes, demonstrating that memory bandwidth can impact solve time by as much as 225% on a simple 8 core test
- Why buying one memory stick and planning to get more later is not the best plan
- CPU scaling in constant vs. increasing memory bandwidth scenarios: our test results
- What to look for in buying a new computer
- An illustrated example using a CFX pump
- How to best configure a new or existing system’s memory for maximum performance, including the windows command to check your memory layout