To reduce the cost of a nightly batch job on your GCP project that uses a large number of VMs and is fault-tolerant, what should you do?