FreeBSD optimizations used by Netflix serving at 800Gb/s - Drew Gallatin - EuroBSDcon 2022
ฝัง
- เผยแพร่เมื่อ 26 ต.ค. 2022
- Quantifying the importance of several FreeBSD optimizations to serving web
traffic at extreme speeds.
In previous talks, I've focused on new optimizations such as NUMA and NIC kTLS offload, which are critical to serve 800Gb/s. In this talk, I'll focus on the
contributions of older, but equally critical optimizations, such as
- TCP segmentation offload (TSO)
- TCP large receive offload (LRO)
- VM optimizations such as the UMA VM pg cache zone and batchqueues
- Asynchronous sendfile
- Unmapped (extpg) mbufs
I will briefly describe the importance of each of these unsung optimizations,
explain why they help, and quantify the value of each optimization using a
metric I call percent-CPU per Gb.
I will wrap up by putting them all together and showing results from an
experimental Netflix server serving production traffic at 720Gb/s.
Now we know why Netflix is always outperforming other streaming services.
Watching this video on my FreeBSD 14.0 CURRENT / xfce 4.18 desktop btw.
As an openBSD enthusiast and plan9 user. Yea Linux is way too fragmented below distro the level no comparison in performance. Nice presentation! Impressive!
It's amazing!!!
Nice Information
Amazing research.
Awesome! Hope it will break another record within next 1 to 3 years.
🎉😮🤘🤤
A lot of content for 1 hr. Just wondering...there is a lot of talk about physical hardware and VM specs. How do you reconsider all of these hardware-specific requirements as a partnership between Netflix and AWS? Does Netflix ask AWS to put these machines in their data centers as a collaborative effort exclusively for Netflix, or what is the situation here? Are these hardware configurations just generic configurations AWS has in place for all AWS users? Thanks
while a FreeBSD user is served at 0kb/s as they can't watch Netflix on their beloved OS.
Thats impressive! 😁
Hmm so FreeBSD is actually good for something