Size an st1 or sc1 volume from its MB/s per TiB baseline and see whether your workload lives inside it or bursts.
st1 and sc1 buy throughput by the terabyte rather than by provisioning: baseline scales at 40 or 12 MB/s per TiB up to a per-volume ceiling, and the burst rate is only sustainable while credits last. Because both types meter I/O in 1 MiB units, small random reads waste most of each metered operation, which is why they suit sequential scans and not databases. Any workload that needs sustained throughput must be sized against the baseline, never the burst figure. Sizing to the burst number is the classic st1 mistake: the volume looks fast in a short benchmark and then falls to a fraction of that rate once the credit bucket empties during the first real batch run.
EBS Throughput
baseline = min(volume ceiling, TiB x MB/s per TiB); size needed = required MB/s ÷ MB/s per TiB; implied IOPS = MB/s x 1024 ÷ I/O size in KiB.
baseline = min(volume ceiling, TiB x MB/s per TiB); size needed = required MB/s ÷ MB/s per TiB; implied IOPS = MB/s x 1024 ÷ I/O size in KiB. st1 and sc1 buy throughput by the terabyte rather than by provisioning: baseline scales at 40 or 12 MB/s per TiB up to a per-volume ceiling, and the burst rate is only sustainable while credits last. Because both types meter I/O in 1 MiB units, small random reads waste most of each metered operation, which is why they suit sequential scans and not databases. Any workload that needs sustained throughput must be sized against the baseline, never the burst figure.
Sizing to the burst number is the classic st1 mistake: the volume looks fast in a short benchmark and then falls to a fraction of that rate once the credit bucket empties during the first real batch run.
This calculator takes 6 inputs: Volume size, Baseline throughput per TiB, Burst throughput per TiB, Per-volume throughput ceiling, Sustained throughput the workload needs, Average I/O size. The pre-filled defaults are a realistic starting point — replace them with figures from your own environment for a result you can act on.
No. st1 and sc1 charge every I/O as a 1 MiB unit and have no provisionable IOPS, so random small-block access performs badly. They are built for log processing, data warehouse scans and big sequential ETL.