HTTP/2 from 2015: Multiplexing, HPACK, and the Limits of Speed
Explore the 2015 HTTP/2 architecture, stream multiplexing, HPACK header compression, and its limits under TCP loss with a realistic web application.
The bottleneck HTTP/2 addressed in 2015
Standardized in RFC 7540 in 2015, HTTP/2 aimed to improve connection efficiency without changing the meaning of HTTP requests and responses. A media-heavy page had to retrieve HTML, style sheets, JavaScript, and numerous small images. HTTP/1.1 had practical concurrency limits on a single connection, so browsers often opened several TCP connections to one origin. That workaround added connection setup, memory use, and network coordination overhead. HTTP/2 instead introduced independent streams carried by one connection. The promise was not that every resource would become smaller or every server faster. Real gains depended on request patterns, network conditions, and the origin's ability to schedule and deliver responses efficiently.
Streams, frames, and a shared connection
HTTP/2 divides messages into binary frames. HEADERS frames carry metadata and DATA frames carry response bodies; each frame belongs to a numbered stream. When a browser requests two resources concurrently, the server can interleave frames over a shared connection. A small stylesheet need not wait until every byte of a large image has been delivered. This flexibility comes with implementation responsibilities: correct stream association, connection and stream flow control, server scheduling, and memory management. Performance reviews should therefore consider time to first byte, resource completion distributions, and concurrency rather than counting connections alone. A single TCP connection still follows the behavior and limitations of its underlying transport.
Why HPACK mattered
Browsers repeat many header fields across requests to the same origin, including cookies, accepted content types, caching metadata, and user-agent information. HTTP/2 uses HPACK, described in RFC 7541, to encode these repetitions more efficiently. Static and dynamic tables allow compact references instead of repeatedly transmitting entire fields. That can meaningfully reduce header overhead for workloads with many small API calls. HPACK does not compress a JPEG or WebP payload; it targets message metadata. Dynamic compression state also introduces operational and security considerations. To evaluate its value, inspect real request traces and distinguish raw header bytes, transferred payload bytes, and any additional processing cost. A faster benchmark in isolation is not enough to establish user-visible benefit.
TCP packet loss still affects every stream
Although HTTP/2 multiplexes streams at the application layer, all streams still share one ordered TCP byte stream. A missing TCP segment can delay delivery of later data until loss recovery completes. Requests that look independent to the browser may therefore experience a shared transport stall. This behavior becomes more noticeable on high-latency or lossy mobile connections. One major motivation for HTTP/3 over QUIC was to improve stream independence when packets are lost. Yet no protocol provides an unconditional speed advantage on every network. Test representative pages under cellular, Wi-Fi, packet-loss, and constrained-bandwidth conditions. Compare repeated measurements rather than relying on a single best-case synthetic benchmark, especially if your audience accesses the site from varied devices and regions.
A measurement plan for a publishing site
Consider a news site with one large hero image, a strip of thumbnails, and related headlines in a side rail. Prioritizing the first visible image, deferring offscreen media, and reducing unnecessary JavaScript often helps regardless of protocol. In browser developer tools, group requests by negotiated protocol and inspect sizes, start times, and completion order. Keep bandwidth and latency controlled for laboratory comparisons, but also look at real-user measurements on mobile devices. Distinguish network transfer efficiency from user experience metrics such as LCP, INP, and CLS. If images are oversized or the main thread is blocked, better connection utilization cannot eliminate those bottlenecks. Measure the entire page before attributing a speed improvement to HTTP/2 alone.
Deployment decisions and the current standard
Evaluate the browser-to-CDN and CDN-to-origin paths separately when enabling HTTP/2. A browser negotiating HTTP/2 with an edge network does not prove that the origin connection uses the same protocol. Check TLS negotiation, ALPN, reverse proxies, and compatibility with older clients. After deployment, compare error rates, connection reuse, p95 latency, and transferred bytes across representative traffic. Remember that the 2015 RFC 7540 specification was superseded by RFC 9113 in 2022, so priority assumptions copied from early tutorials may not match current implementations. The goal is not to display the latest protocol label in a dashboard. It is to deliver measurable improvements in load time, stability, and resource efficiency for actual visitors.
Comments
No comments yet. Be the first to contribute.
Reading “2015'ten bugüne HTTP/2: Çoklu akış, HPACK ve gerçek hızın sınırları” raised this question for me: a small reproducible performance comparison would help.
Editor note: Results can vary by scale. Cost, latency and user experience should be compared together.
The learning curve and accessibility also matter for actual users.
A practical case study with load testing or a pilot would help establish the trade-offs.
✎ Write a comment