NYCBSDCon 2014: Serving one-third of the Internet via FreeBSD
BSDTVCited 5× on HN, 2014-2019
14K views · 12 years ago · talk · dated
Scott Long of Netflix presents the architecture of the OpenConnect content delivery network at NYCBSDCon 2014. He explains why Netflix runs FreeBSD 10 on its appliances and what the future holds.
Cited as the primary source for why Netflix uses FreeBSD and UFS rather than ZFS on its CDN boxes, and for its GPL reasoning.
45:42Why did Netflix use NGINX and FreeBSD to build their own CDN? by Gleb Smirnoff
NGINX
40K views • 11 years ago
40:20Why We Run FreeBSD current at Netflix – Drew Gallatin
OpenFest Bulgaria
21K views • 2 years ago
47:42Serving Netflix Video at 400Gbps on FreeBSD by Drew Gallatin
EuroBSDCon
9K views • 5 years ago
44:05FreeBSD optimizations used by Netflix serving at 800Gb/s - Drew Gallatin - EuroBSDcon 2022
EuroBSDCon
28K views • 4 years ago
18:51IPFS Alpha Demo
Juan Benet
272K views • 11 years ago
1:21:16A New Way to look at Networking
Google
52K views • 19 years ago
1:21:16A New Way to look at Networking
GoogleTalksArchive
26K views • 14 years ago
1:21:20A New Way to look at Networking
Google TechTalks
30K views • 19 years ago
32:30Benno Rice: The Tragedy of systemd -- BSDCan 2018
BSDCan
25K views • 8 years ago
7:06Putting Google's Network to work for You
Google for Developers
23K views • 12 years ago
47:18OpenFlow @ Google - Urs Hoelzle, Google
Open Networking Summit
47K views • 14 years ago
1:59:29Tutorial: New and Updated - Everything You Always Wanted to Know About Optical Networking
NANOG
58K views • 9 years ago
35:31June 2021 FreeBSD Developer Summit: Raid-Z Expansion
FreeBSD
4K views • 5 years ago
47:18The Tragedy of systemd
linux.conf.au
1.3M views • 7 years ago
56:57IPv6 Summit 2014: Wells Fargo adopting IPv6
Rocky Mountain IPv6 Taskforce
237 views • 11 years ago
28:21T-Mobile's journey to IPv6
NANOG
1K views • 8 years ago
46:23High Volume, Bandwidth Heavy Infrastructure On The Cheap - Kyle Drake
PDX Node
1K views • 9 years ago
1:11:44DebConf 14: QA with Linus Torvalds
Adnan Hodzic
737K views • 12 years ago
53:14Mastering Chaos - A Netflix Guide to Microservices
InfoQ
2.3M views • 9 years ago
48:52Jordan Hubbard - FreeBSD: The Next 10 Years
iX
24K views • 11 years ago
5 comments from Hacker News
Pretty sure it's still UFS+J. They've mentioned in the past that the benefits of ZFS don't really apply to the CDN boxes: if a drive dies, it's simply removed from the local cache, and files are re-downloaded/redistributed as necessary across the remaining cache drives the next time the appliance refreshes from its upstream origin. Additionally, ZFS prefetching was actually less performant in some cases from what I remember -- or at least providing no benefit -- because of how it couldn't really anticipate the I/O patterns of video segment requests from users. Additional reference: - link ("…but caching is futile.") - ▶ this video
Netflix delivers all video streaming content via FreeBSD which makes up 37% of all internet traffic. link They also contribute to FreeBSD core. It is not just an "appliance" as some people seem to want to point to. link link ▶ this video In addition, a guy I met, who worked at Yahoo, told me about 25% of Yahoo still runs FreeBSD as of two years ago.
Scott Long has summarized it a few times, eg, from link : • Availability of expertise, outstanding community • Works well, good vendor support • No GPL • Features used: – SUJ – gmirror – boot drive only – AIO – Dtrace, HWPMC – TCP Stack, modular CC It's missing "Integrates tightly with nginx: Async I/O and sendfile()" which Scott included in ▶ this video . In the video he elaborates a bit more about the GPL... edit: formatting
Actually it turns out they don't really have a need for most of the features of ZFS. So they don't use it. ▶ this video
> I'm assuming whatever file system is on the disks... Netflix uses UFS+J on FreeBSD 10. Here are some notes from their talk at NYCBSDCon 2014: - 400,000 stream files per appliance. - 5,000 - 25,000 client streams per appliance. - 300 - 500 streams coming off each disk all the time. - Attempt to buffer 1MB ahead, but caching is futile. - Result is completely random disk workload. - System becomes limited by disk latency and CPU load. Video here: ▶ this video