Evaluate the Impact of Hyper-threading

All the simulation runs discussed so far were executed with hyper-threading enabled, meaning computational resources were configured with threads_per_core=2. This is the default setting for virtual machines on Inductiva (learn more here).

In traditional HPC environments, however, it's common practice to run one thread per physical core, with hyper-threading disabled. This approach helps avoid resource contention and can lead to more predictable and consistent performance.

To disable hyper-threading and ensure only physical cores are used, simply configure your MachineGroup with threads_per_core=1:

cloud_machine = inductiva.resources.MachineGroup( \
    provider="GCP",
    machine_type="c4d-highcpu-32",
    threads_per_core=1,
    data_disk_gb=20,
    spot=True)

The number of available vCPUs will be halved, but the underlying number of physical cores remain the same.

Performance with Hyper-threading Disabled

We repeated the S2_f5/polynya2D_20211007 scenario tests on the same machine types used before, but this time with hyper-threading disabled. Here are the execution times and associated costs:

Comparing Results: With vs. Without Hyperthreading

From these results, we observe:

• At lower vCPU counts, the difference in performance between hyper-threaded and non-hyper-threaded runs is minimal.
• At higher vCPU counts, however, disabling hyper-threading leads to better performance and lower cost per hour.

This trend suggests that contention for RAM access becomes a limiting factor when too many threads compete for shared hardware resources.