Performance
Last updated 17 November 2026
A Volume reads and writes on the local disk of the server its Machine runs on, so while the Machine is awake it is as fast as a local disk, up to the disk limit of the Machine’s shape. A larger shape raises that limit.
Each shape’s disk limit
| Shape | vCPUs | Reads, and writes, each a second | Bytes read, and written, each a second |
|---|---|---|---|
s-1 | 1 | 2,500 | 50 MiB |
s-2 | 2 | 5,000 | 100 MiB |
m-4 | 4 | 10,000 | 200 MiB |
m-8 | 8 | 20,000 | 400 MiB |
b-16 | 16 | 40,000 | 800 MiB |
b-32 | 32 | 40,000 | 800 MiB |
Reads and writes each have their own limit. Each limit is shared by all of the Machine’s disks, its own disk, its scratch disk and its Volume, so a busy scratch disk leaves less for the Volume. These are ceilings, not what every Machine gets: a Volume answers as fast as the server’s disk allows, up to them. Tonbo has not published measured Volume speeds. The rest of what a shape gives is on Limits.
See when a Machine reaches its limit
A Machine’s metrics show volume_throttled_requests and disk_stalled_seconds above zero when it is held to its shape’s disk limit (see Watch how it is used). A Volume’s own metrics show the bytes and requests its Machine read and wrote, and how long the Volume took to answer them, the median and the 99th percentile (see Watch what your Volumes store).
The first reads after a wake on another server
When a Machine wakes on another server than the one that holds its Volume’s working copy, the Volume is ready at once, but each block your program reads for the first time is fetched from the region’s storage before it is answered, while the rest fills in behind. So those first reads are slower than a local disk, and reads become local-disk fast as the copy fills in (see A wake on the same server, or on another). A wake on the same server reads at local speed from the start.
When fsync waits
fsync returns once a write is on the server’s disk, so it is quick, except when copying to the region’s storage has fallen five minutes behind: it then waits until the copy catches up (see Writes are held).