Dropbox recently experimented with HTTP/3 to improve network latency. Harnessing the enhanced head-of-line blocking in HTTP/3, the team at Dropbox observed a notable reduction in latency, particularly at the 90th percentile (p90) and higher.
Tiffany Fong, Mike Lyons, and Nikita Shirokov from the Retrieval Experiences and Traffic team at Dropbox described the experiment in a blog post. Network latency is influenced by various factors such as the time of day, local network conditions, and the distance between the user's location and Dropbox server locations. The team found that network latencies in Europe were twice as high as those in North America, while in Asia, latencies were three times higher compared to North America.
The team set up the experiment by designing a test site that made specific API requests over HTTP/3 without impacting the users of the main site. For two weeks from December 2022 to January 2023, the team triggered 300,000 requests per day. To simulate real-world scenarios at Dropbox, the requests were executed in parallel, emulating concurrent actions. The HTTP/3 tests were performed once for each page load and after the completion of a user's search.
HTTP/3, the third significant iteration of the Hypertext Transfer Protocol, was officially published as a Proposed Standard in RFC 9114 by the Internet Engineering Task Force (IETF) on June 6, 2022.
To replicate real-world scenarios, the team at Dropbox implemented a systematic approach. They initiated a pre-warming process by triggering two sequential HTTP/3 requests to populate the cache. Subsequently, they attempted to utilize HTTP/3 for all subsequent connections following the initial HTTP/2 connection. Next, they executed five simultaneous HTTP/2 requests to a no-op API endpoint, recording the network time for each request. Finally, in the last stage, another set of five parallel requests was made to the same no-op API endpoint using HTTP/3, and the elapsed network time was logged for analysis.
Using HTTP/3 caused a latency reduction of 48ms (or 13%) at p90 and a reduction of 146ms (21%) at p95. The team concluded that HTTP/3 is better at handling packet drops in parallel connection by eliminating head-of-line blocking. The impact of HTTP/3 on network latencies became clearer, particularly when analyzing specific regions at higher percentiles. In the Asia region, HTTP/3 demonstrated a significant reduction in network latencies of approximately 77ms (p90) and a remarkable 200ms (p95).
Source: Investigating the impact of HTTP3 on network latency for search - Dropbox
The experiment caught the attention and piqued the curiosity of the tech community on YCombinator. One of the users clementmas commented, "Can we use HTTP3 today? Is it widely supported?" The community answers affirmed the deployment of HTTP/3 with Google and Facebook, while Nginx 1.25 has HTTP/3 in its roadmap.
In conclusion, the Dropbox team has gained a deeper understanding of the advantages offered by HTTP/3, building upon their existing theoretical knowledge. Following the successful experiment, they now have plans to implement HTTP/3 in their production environment.