Spot instances for HPC workload: use cases


Has anyone ever used spot instances for HPC workload? Do we have any public reference use-case?

IFAIK, Spot can be used if the job can checkpoint his status (so we can restart in case of a termination). Otherwise if the job runs for many hours, you run the risk to have your job terminated before it’s completed.

Looking forward to your comments.

I'll appreciate if you can share links to public use-cases or guides.

asked 7 years ago377 views
2 Answers

See the TLG Aerospace use case study for tightly-coupled:

Fermi labs for HTC:

Spot is often used for both high throughput computing HPC and tightly-coupled HPC. This is because most HPC workloads are short lived and already have check pointing. Long time HPC users are used to less reliable environments or forced check pointing to get off a supercomputer within a certain number of hours.

While a lot of HPC is run on spot, unless the job completes within two minutes of notice, many users are OK with losing a job. Those that aren't, can checkpoint every so often, and relaunch from the EBS volume. Still, while users could recover most jobs, they often don't bother to try to save a case. They check spot prices beforehand, bid at a higher value, and rerun the job if the spot price is exceeded. This is because relatively few jobs are lost when a bit of care is taken, the savings are already large, and relaunching a lost job is easy.

answered 7 years ago
Accepted Answer

This blog post describes how to use Spot instances for CAE workload such as LS-DYNA using checkpointing:

answered a year ago

You are not logged in. Log in to post an answer.

A good answer clearly answers the question and provides constructive feedback and encourages professional growth in the question asker.

Guidelines for Answering Questions