Latency and Capacity Numbers
The numbers worth having in your head for back-of-envelope reasoning about where time and bandwidth go.
Round-trip latency by distance
| Path | Typical RTT |
|---|---|
| Same rack / same datacenter | ~0.5 ms |
| Same metro region (multi-AZ) | ~1-2 ms |
| Cross-country (e.g. US coast-to-coast) | ~40-60 ms |
| Intercontinental (e.g. US to Europe or Asia) | ~150-300 ms |
| Geostationary satellite | ~600 ms+ |
- L1 cache reference: ~1 ns
- Main memory reference: ~100 ns (100x an L1 hit)
- SSD random read: ~100 microseconds (1,000x memory)
- Spinning disk seek: ~1-10 ms (10,000x+ memory)
- Same-datacenter round trip: ~0.5 ms
- Cross-country round trip: ~40-60 ms
- Every hop up this ladder is roughly one to two orders of magnitude slower than the one before it - a cache miss that turns into a network call can cost 100,000x more than the cache hit would have
Link bandwidth
| Link | Typical bandwidth |
|---|---|
| Wi-Fi (802.11ac/ax, real-world) | ~100 Mbps-1 Gbps |
| Gigabit LAN (wired) | ~1 Gbps |
| Typical home/office broadband | ~100 Mbps-1 Gbps |
| Datacenter backbone / core links | ~100 Gbps-400 Gbps+ |
A higher-bandwidth link moves more bytes per second once data is flowing, but it does nothing for the time a single round trip takes. A satellite link can have plenty of bandwidth and still feel sluggish for chatty request/response protocols, because every round trip pays the ~600 ms propagation delay regardless of how wide the pipe is.
MTU, MSS, and jumbo frames
| Term | Typical size | Notes |
|---|---|---|
| MTU (standard Ethernet) | 1500 bytes | Largest frame payload a link will carry unfragmented |
| MSS (over standard Ethernet) | ~1460 bytes | MTU minus IP header (20) and TCP header (20) |
| Jumbo frames | 9000 bytes | Datacenter NICs only; both ends and every hop in between must support them |