I finally got AWS ParallelCluster working (after a long, painful battle). I keep referencing some heterogeneous AWS things I've written down in various places so I figured I should put them all in one place so I can find it more easily later.

GPU Notes

The CUDA architecture corresponding to each gencode:

Alternatively, for some GPUs which are available through AWS:

Instance Types

Use this link to get the most up-to-date prices for each instance type. The prices below are from a particular reference day.

For reference, there are 720 hours in a month and 8760 hours in a year, so:

• $**0.10** / hour translates to$2.40 / day, $**72** / month, or$872 / year.
• $**0.50** / hour translates to$12.00 / day, $**360** / month, or$4,380 / year.
• $**1.00** / hour translates to$24.00 / day, $**720** / month, or$8,760 / year.
• $**2.00** / hour translates to$48.00 / day, $**1,440** / month, or$17,520 / year.
• $**4.00** / hour translates to$96.00 / day, $**2,880** / month, or$35,040 / year.

Other non-GPU instance types:

Availability Zones

Use this link to get availability zones for each region. The zones below are for us-east-1.

ParallelCluster

• There's some weird issues with some of the instance types and elastic IPs which cause the cluster to hang while being created. Can potentially fix by adding ElasticIp: true under HeadNode.Networking and AssignPublicIp: true under Scheduling.SlurmQueues.Networking in the config.
• To use A100s (i..e, p4d.24xlarge), you need to stop the subnet from automatically assigning a public IP. To do so:
  • Go to the VPC console here
  • Click on "Subnets" in the left sidebar.
  • Select the subnet you want to use.
  • Click on "Modify auto-assign IP settings" in the "Actions" dropdown.
  • Uncheck "Auto-assign IPv4" and click "Save".
• See the full cluster configuration here or just the required options here

Important Logs

• /var/log/parallelcluster/slurm_resume.log - Logs for the slurm resume script
• /var/log/parallelcluster/clustermgtd - Logs for the cluster manager daemon (requires root access)
• /var/log/slurmctld.log - Logs for the slurm controller (requires root access)

FSx Lustre

• Console
• FSx mounts are restricted to a single availability zone

Aurora RDS

The dev and prod instance types for Aurora that AWS recommends are:

• Prod
  • db.r6g.2xlarge
    • 8 vCPUs
    • 64 GiB RAM
    • $1.038 per hour or$174.38 per week
  • This is under the "memory optimized" category
• Dev
  • db.t4g.large
    • 2 vCPUs
    • 8 GiB RAM
    • $0.146 per hour or$24.53 per week
  • This is under the "burstable" category

Here are some additional links which I found helpful:

• Connecting to an RDS database