Faster first boot on large instances

This applies to AMIs that carry a large model on disk, such as the Qwen3.8-27B server. The first start of a new instance is slow because of how AWS restores a disk from a snapshot, not because of the model server. Later restarts are much faster.

What we measured

Qwen3.8-27B server AMI, cold launch of a new on-demand g6e.xlarge in us-east-1: 30 minutes 29 seconds from launch until the server answered its health check, and 30 minutes 32 seconds until the first authenticated answer. During that time the disk delivered about 12 to 22 MB/s. Once the data was on the disk, the model itself loaded and compiled in a few minutes. These are single measurements, not averages.

Why

An EBS volume created from a snapshot loads each block from S3 the first time it is read. A model server that reads tens of gigabytes at start-up therefore waits on that first read. Reading the same blocks again, including after a stop and start of the same instance, is fast. Adding gp3 throughput or IOPS to the volume does not remove this wait.

What you can do

1. Plan for it. Treat the first start of an instance as taking about half an hour for the 27B AMI. Do not put a new instance behind a load balancer or an auto-scaling policy that expects it ready in a few minutes.
2. Stop and start instead of terminate and relaunch. A stopped instance keeps its volume, so the blocks already read stay local. Restarting the model server on a running instance took a few minutes in our tests. You pay for the volume while it is stopped.
3. Start earlier. Launch the instance before you need it, and wait for the health check at /health to return before sending traffic.
4. Set a volume initialization rate (not yet measured by us). AWS can load a snapshot-restored volume at a rate you choose, from 100 to 300 MiB/s, for an additional charge. It cannot be stored in the AMI, so a plain launch from the Marketplace console does not use it. You can set it when you launch through the EC2 API, a launch template, or the CloudFormation AWS::EC2::LaunchTemplate and AWS::EC2::Volume resources. It is not available on AWS::EC2::Instance. We have confirmed AWS accepts the setting at launch; we have not yet measured how much faster the model starts, so we do not quote a time for it.

What does not work

Fast snapshot restore is enabled on the snapshot owner's account and does not carry over to the copy of the snapshot that your account receives from the Marketplace. Reading the model files in parallel did not make the disk faster in our tests.

Questions or a measurement that differs from ours: contact us. See also launching an instance.