Running Acoustic Simulations on a Lambda Cloud Server
-
The data sheet shows the simulations can be run on a cloud server. I am attempting to run a simulation on a Lambda Labs Cloud server with an A10 NVIDIA GPU. I am connected with SSH to the cloud GPU server. The network also appears when I scan for available servers. However, the log shows "Error: unable to connect to remote Ares [150.136.59.24:22]". How do I configure the network server to run jobs on a Cloud GPU? Must I use port 18701? Must I install Sim4Life on the Cloud Server and run Sim4life as an Ares server?
-
I eventually configured Sim4Life on a Cloup GPU workstation (paperspace.com). The workstation has an Intel Xeon E5-2630 v3 processor, 16 virtual cores, 90 GB of RAM, and two NVIDIA Ampere A6000 GPUs, each with 48 GB of GPU memory. For Exablate Neuro transducers (220 and 670 kHz) simulations, the workstation improved simulation speed from about 200 Mcell/s with an 8-core OpenMP configuration, to about 12000 Mcells/s with two GPUs.
The workstation allows simulations (1449x1149x816 grid, 200 periods, 670 kHz, 0.22 mm max step for 10 points per wavelength) in about 40 minutes with Fourier-domain sensor recordings and about 4 hours with Time-domain sensor recordings. Time-domain simulations seem only compatible with a single GPU, while Fourier-domain sensor simulations can use both GPUs. Also, the Fourier domain field sensor recordings create simulations that look unstable (or less accurate) compared to the time-domain field sensor recordings, with all other factors equal. Only the time-domain simulations have readily matched previous studies, but the simulation time is still too long.
- Can the time-domain sensor acoustic simulation be run on multiple GPUs?
- Also, most of the time for Fourier-domain simulations is spent "allocating memory for the voxel array." Is there a way to configure the GPU to preallocate memory or another way to reduce the amount of time to distribute the array memory?