Performance Tuning EC2 Instances

[vinnyglennon]

http://www.slideshare.net/brendangregg/performance-tuning-ec2-instances

[helper]

The video that goes with these slides is here: https://www.youtube.com/watch?v=7Cyd22kOqWc

[jordanthoms]

Can someone comment on why they use FreeBSD for content delivery (vs linux as a default choice)? Seems people use BSDs for certain high-performance tasks and I'm curious what advantages it has.

[brendangregg → jordanthoms]

Scott Long has summarized it a few times, eg, from http://people.freebsd.org/~scottl/Netflix-BSDCan-20130515.pd...:

  • 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 http://youtu.be/FL5U4wr86L4?t=18m30s .

In the video he elaborates a bit more about the GPL...

edit: formatting

[cbsmith → brendangregg]

Wouldn't you say that Linux integrates just as nicely with nginx's async I/O and sendfile support?

[m0th87 → cbsmith]

I thought *BSDs used kqueue, which allegedly had worse performance at high concurrency vs Linux's epoll?

EDIT: This seems to suggest kqueue may have better performance due to less syscall overhead: http://www.eecs.berkeley.edu/~sangjin/2012/12/21/epoll-vs-kq...

[cbsmith → m0th87]

Yeah, it's the opposite, though there are other mechanisms that could do better than epoll, but they need more refinement.

[donavanm → cbsmith]

Actually o_async is, imo, broken on linux. Forcing it to direct precludes the use of readahead or the page cache.

[cbsmith → donavanm]

Yeah, O_ASYNC on Linux kinda sucks.

[dbarlett → jordanthoms]

As of 2012, apparently IPv6 support/performance: http://lists.freebsd.org/pipermail/freebsd-stable/2012-June/...

[lazyant]

"Swappiness is set to zero to disable swapping..." for many (most, older?) distros setting swappiness to zero won't disable swapping completely.

[nwmcsween]

vm.swappiness != 0 is useful if you use something like zswap.