Version 2.0.1 is a bug-fix release of the 2.0 series. Existing users can upgrade to the new version to avoid the issues listed in the changeling.
Version 2.1.0-beta1 is the first beta release of 2.1, with a new throttle-response handler for per-response bandwidth throttling, and an enhancement to the status handler (pull #893). It also includes two new features that improve HTTP/2 performance: TCP latency optimization and support for
link: rel=preload headers in informational response (pull #916).
With TCP latency optimization, users can expect 1 RTT or more reduction in time-to-render if the main resource (i.e. HTML) is much larger INITCWND (typically ~15KB).
The reduction comes from the fact that with the optimization enabled, H2O tries to keep the amount of HTTP/2 frames kept unsent in the TCP send buffer very small (to just two packets) during the slow-start phase. Since the amount of data unsent is kept small, the server can switch to sending a resource that blocks the rendering path (e.g. CSS) immediately when it receives a request for such resource, instead of pushing the HTML body stored in the TCP send buffer. As CWND grows, the connection handling switches to bandwidth-optimization mode, that pre-fills more data into the send buffer so that the kernel can send additional data immediately after receiving ACKs without user-space intervention.
link: rel=preload headers in informational response helps web developers utilize HTTP/2 push. Use of the link header is becoming the standard way to instruct HTTP/2 servers to start pushing assets. The downside of the approach is that application servers typically cannot send the header until it generates the final response. Generation of the final response often involves time-consuming operations such as access to the database, keeping the HTTP/2 connection idle for that period.
Use of informational response lets us use the time slot for pushing asset files. Application servers can now send an informational response with link rel=preload headers to H2O to start pushing the asset files, then perform heavy tasks, and send the final response. Use of 1xx response will not cause interoperability issues, since only the final response is sent to the client connected to H2O.
Details of the two optimizations were covered in my presentation at Tokyo RubyKaigi 11. The slides are shown below: