“100 milliseconds” is not a performance guarantee for every region and workload. It is a useful budget for breaking down the beginning of a request. Actual times vary, but diagnosis usually starts with four stages:
- Resolution. If the recursive resolver detours to a distant authoritative server, you’ve lost tens of milliseconds on the very first hop.
- The handshake. Every extra TLS round trip costs more the farther you are.
- Fetching. An image or JS bundle that misses cache must wait for the origin connection and origin processing.
- Application first byte. For AI requests, separate gateway time from the model’s queue and time to first token.
One total duration says that a request is slow. Stage-level timing says whether to work on DNS, transport, cache or the application.
Latency stages controlled by the edge
After a domain is connected to IXCDN, each stage has a corresponding control:
- Nearest resolution. Anycast + GeoDNS smart resolution read the visitor’s network line or geography and resolve them to the nearest suitable edge ingress — carriers and overseas regions each get a different answer, with unmatched visitors falling back to the default line, instead of everyone detouring to one distant authoritative server.
- Connection setup. HTTP/3 (QUIC) and TLS 1.3 are used when the client and route support them, reducing avoidable round trips.
- Cache and origin. Cacheable content returns from the edge on a hit. Misses reuse regional origin connections, while Origin Shield can collapse concurrent fetches for the same object.
- AI ingress. The AI gateway provides an OpenAI-compatible interface. Model queueing and generation still belong to the upstream and must not be reported as network latency.
These controls share one delivery path, but their effect should be checked with before-and-after stage data, not a single total-duration sample.
GeoDNS: responses selected by network path
One A record does not always need to return the same answer to every visitor. IXCDN smart resolution can select different ingress points by carrier and region, use priority when several rules match, and fall back to a default line. “Nearest” means a suitable route based on network reachability and node health, not map distance alone.
Regional origin fetch: shorten the edge-to-origin path
On a cache miss, the origin hop may dominate latency. Regional fetch reuses connections by region, and Origin Shield can combine concurrent requests for one object. Origin location still matters; these mechanisms reduce repeated connections and repeated fetches rather than removing physical distance.
Performance metrics and billing units
IXCDN meters delivered CDN traffic, model tokens and compute separately. This lets a team compare cache hit rate and origin traffic with the corresponding cost instead of relying on one blended number. Current free allowances and production prices are the values shown in Console.
Before-and-after validation with logs
Logs and real-time analysis should answer three concrete questions: which ingress handled the request, whether it hit cache, and how much time was spent on origin or upstream work. AI traffic also needs separate first-token and full-response measurements. Compare the same workload under similar conditions before calling a change an improvement.
Low-risk hostname rollout
Connect a test hostname or low-risk service first and record a baseline. Then check DNS, TLS, cache status and logs before expanding traffic. That sequence produces better evidence than a one-off speed test.