Reference
Numbers to estimate with. SI units: 1 KB is 1,000 bytes.
Latency numbers
- L1 cache reference0.001 µs
- Branch mispredict0.003 µs
- L2 cache reference0.004 µs
- Mutex lock/unlock0.017 µs
- Main memory reference0.1 µs
- Compress 1 KB with a fast codec2 µs
- Send 1 KB over a 1 Gbps link8 µs
- Read 4 KB randomly from an NVMe SSD20 µs
- Read 1 MB sequentially from memory50 µs
- Read 1 MB sequentially from an NVMe SSD300 µs
- Round trip within a datacenter500 µs
- Disk seek (spinning disk)5,000 µs
- Read 1 MB sequentially from a spinning disk5,000 µs
- Round trip US East ↔ US West70,000 µs
- Round trip California ↔ Netherlands150,000 µs
Log scale: each gridline step is 10×.
Sources: Peter Norvig and Jeff Dean, approximate timing for various operations; Colin Scott, Latency numbers every programmer should know (interactive). Rounded; real systems vary by 2–10×.
Nines to downtime
%
- Per year
- 8.77 h
- Per 30 days
- 43.2 min
- Per week
- 10.1 min
- Per day
- 1.44 min
- 99%5,259.6 min
- 99.9%525.96 min
- 99.95%262.98 min
- 99.99%52.6 min
- 99.999%5.26 min
Log scale: each gridline step is 10×.
Load, storage and bandwidth
×
bytes
%
days
copies
- Requests / day
- 200,000,000
- Average load
- 2,315 rps
- Peak load
- 6,944 rps
- Peak bandwidth
- 111.1 Mbps
- New data / day
- 40 GB
- Stored, 365 days
- 43.8 TB