Hacker News Reader

Comment by SEJeff | Hacker News Reader

Linux has been beating BSD for at least 8-10 years when it comes to TCP. When it comes to new features in TCP-land, Linux easily beats it. Google added Receive Side Scaling / Receive Flow Steering to Linux years ago, and it is still a WIP in FreeBSD as an example. Also take a look at much of the bufferbloat research recently that has been merged into Linux, etc.

View on HN

The RSS is Linux is not particularly useful (for netflix scale workloads) because it does not integrate the RSS hashing across the entire stack. So all you get is connection sharding. With real RSS, as done in Windows and FreeBSD where the kernel has intimate knowledge of the hash key and algorithm, you can use RSS to split the TCP hash table up and make it per-CPU. By using multiple accept sockets for per-CPU workers, you can effectively keep everything for a single connection on a CPU, and run almost everything with no cross-CPU contention. You can't move the connection around at will between CPUs, but you don't care to, because no connection is special (in a netflix workload), it just one of tens of thousands.

Adrian Chadd did most of the FreeBSD RSS work, and gave a good talk about it at BAFUG: https://www.youtube.com/watch?v=7CvIztTz-RQ

The RSS in Linux was just used for load spreading (the last I checked, I haven't used Linux much since I left Google 1.5 years ago). If this has improved, I'd love to hear about it.

Linux RFS depends on the packets being dispatched to the correct CPU for the connection by the interrupt handler running wherever the packet happened to land. This has cache & memory locality implications, especially on NUMA.

Linux aRFS lets the NIC do the steering. Unfortunately, each connection requires an interaction with the NIC to poke it into the steering table, and most NICs can't steer 100,000 connections.

So, to sum up, Linux has a lot of cool tech for steering individual connections and support for that varies greatly by NIC. Windows and FreeBSD use standard RSS to predictably steer an unlimited number of connections. For a large CDN server, the latter is more useful. However, for low-latency / high bandwidth applications, I can see the advantage to aRFS.

I wouldn't say 8 to 10 years. A major bug in Linux's default congestion control algorithm was only fixed just last year:

https://bitsup.blogspot.com.au/2015/09/thanks-google-tcp-tea...

Linux is the platform of choice for bufferbloat research, although FreeBSD isn't far behind in adopting the results of it:

https://lists.freebsd.org/pipermail/freebsd-ipfw/2016-April/...

I guess my information was not just outdated, but clearly wrong.

Don't get me wrong, BSD is still absolutely solid, but for anything cutting edge, Linux is spanking the pants off of it.

This is an adolescent evaluation. FreeBSD will have a new TCP stack with BBR made public in a couple months. It will be easier to correctly deploy and more cohesive than Linux. The entire packet path is more cohesive and easier to debug and tune using DTrace although Linux might have caught up here recently. By volume, FreeBSD is doing at least 30% of Internet facing traffic between a well known company and some quieter giants. BTW it only took a half dozen people collaborating between 3 companies about 2 years catch up to state of the art.

I've done a great deal of reading and research on OS ethos, IMO a thriving and production worthy operating system can be maintained with as few as 40 people in total. The superiority of Linux feels exaggerated, and systems innovation has chilled because of it.

As is often the case it depends on the specifics of the application and on those building a solution. As far as raw performance is concerned FreeBSD performs very well.

Netflix gets nearly 100Gbps from storage out the network on their FreeBSD+NGINX OCA appliances. Some details in the "Mellanox CDN Reference Architecture" whitepaper at http://www.mellanox.com/related-docs/solutions/cdn_ref_arch..... The closest equivalent I've found on Linux was a blog post on BBC streaming getting about 1/4 of the performance.

Chelsio has a demo video (with terrible music) using TCP zero copy of 100Gbps on a single TCP session, with <1% CPU usage https://www.youtube.com/watch?v=NKTApBf8Oko.

At SC16 NASA had a "Building Cost-Effective 100-Gbps Firewalls for HPC" demo, using FreeBSD and netmap: https://www.nas.nasa.gov/SC16/demos/demo9.html

You're spot on regarding the app and FreeBSD performing very well. Don't disagree with you one bit. Also, great link on the Netflix CDN work, they're doing some really fascinating stuff. It is nice to see the openness.

I work directly with both of the gents who gave this talk about 100G networking[1] (on Linux) and still find that much of the actual cutting edge research is done on Linux. Perhaps I'm biased! I've also been to one of Mellanox's engineering offices (Tel Aviv) to speak with their engineers at my previous employer 7-8 years ago. They told me they do most all of their prototyping and initial development on Linux, and RHEL to be specific. Then then port to other platforms.

Maybe I was wrong on some of this, but my use case (due to my employer's industry being finance) is lower latency, where Linux absolutely and positively crushes anything else.

    [1] http://events.linuxfoundation.org/sites/events/files/slides/100G%20Networking%20Toronto_0.pdf

Mellanox is now one of the role model vendors in the FreeBSD ecosystem. They have a handful of BSD developers as well as sales and support staff that are in tune with the needs of high scalability FreeBSD users.

Maybe I was wrong on some of this, but my use case (due to my employer's industry being finance) is lower latency, where Linux absolutely and positively crushes anything else.

Actually, while we're on the subject, SmartOS with CPU bursting from illumos is the leader in low latency trading:

http://containersummit.io/events/sf-2015/videos/wolf-of-what...

That is a slick platform they've built, but I still don't see how it is competitive with Linux for very low latencies. He mentions trading at microseconds, but we're building microwave radio networks to trade at nanoseconds. Unless this has changed extremely recently, Solaris/Illumos and hence SmartOS still don't have tickless kernels. I recall Solaris having a 100hz tick by default which you could change to 1000hz with a special boot flag. Linux has had dynticks since fairly early 2.6 kernels and with the modern 3.x kernels (RHEL7+), you've got full on tickless via the nohz_full options. Without this, the kernel interrupts the application to use cpu time.

Additionally, I don't believe (Experts please correct me if this is wrong) SmartOS has an equivalent to Linux's isolcpus boot command line flag (or cpu_exclusive=1 if you're in a cpuset) to remove a cpu core entirely from the global scheduler domain. This prevents any tasks from running on that CPU, including kernel threads. Kernel threads will still occasionally interrupt applications if you simply set the affinity on pid 1 so that does't count.

These two features, along with hardware that is configured to not throw SMIs, allow Linux to get out of the way of applications for truly low latency. As far as I'm aware, this is impossible to do in Solaris/SmartOS. I'm not even getting into the SLUB memory allocator being better or the lazy TLB in Linux massively lowering TLB shootdowns, etc, etc. There is a reason why virtually every single major financial exchange in the world runs Linux (CME in Chicago, NYSE/NYMEX in New York, LSE in London, and Xetra in Frankfurt), it is better for the low latency use case.

You asked for an expert to correct you if you're wrong, so here it is: this is just completely wrong and entirely ignorant of both the capacity of the system and its history.

On timers: we (I) added arbitrary resolution interval timers to the operating system in 1999[1] -- predating Linux by years. (We have had CPU binding and processor sets for even longer.) The operating system was and is being used in many real-time capacities (in both the financial and defense sectors in particular) -- and before "every single major financial exchange" was running Linux, many of them were running Solaris.

[1] https://github.com/joyent/illumos-joyent/blob/master/usr/src...

Thank you Bryan for the correction, I did after all ask for it :)

One final question while I've got you that your response didn't seemingly address. Does the cyclic subsystem allow turning off the cpu timer entirely ala Linux's nohz_full? If so, I stand corrected.

Yes, it does -- the cyclic subsystem will only fire on CPUs that have a cyclic scheduled, which won't be any CPU that is engaged in interrupt sheltering via psradm.[1] This is how it is able to achieve hard real-time latency (and indeed, was used for some hardware-in-the-loop flight simulator systems in the defense sector that had very tight latency tolerence).

[1] https://illumos.org/man/1m/psradm

You should really chat to some HFT folk in NYC before making that conclusion.

Thanks for the reference to the whitepaper. FWIW, I'm the kernel guy at Netflix who has been doing the performance work to get us to 100Gb/s from a single socket.

Another interesting optimization we've done (and which needs to be upstreamed) is TLS sendfile. There is a tech blog about this at http://techblog.netflix.com/2016/08/protecting-netflix-viewi.... We don't have a paper yet about the latest work, but we're doing more than 80Gb/s of 100% TLS encrypted traffic from a single socket Xeon with no hardware encryption offloads.

I just wanted to thank you publicly for all your hard work on this. The community will benefit greatly from this. If I recall, correct me if I am wrong, didn't you also port FreeBSD to the Alpha many moons ago? I loved the Alpha and it broke my heart when it died. Sad panda :(

Doug Rabson did most of the early work on alpha. I sent him enough patches that he sponsored me for a commit bit. My primary desktop for several years was running FreeBSD/alpha. First was an AlphaStation 600, then an API UP1000.

I was very sad when alpha got axed, but I agreed with killing it. FreeBSD is about current hardware.

What's WIP about FreeBSD's RSS?

Reply on news.ycombinator.com