Announcement

Collapse
No announcement yet.

Netflix Optimized FreeBSD's Network Stack More Than Doubled AMD EPYC Performance

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • #21
    Originally posted by Etherman View Post
    Secret source is an possibility on linux too, so what software license war are you talking about?
    It depends. In this case, they touch NUMA and the in-kernel network stack. On Linux both of them would clearly be derivative work, and covered by the GPL, so no secret source. They could circumvent this by using DPDK instead, but if they need to share the NICs between several processes (I'd expect them to be dedicated for those tasks) they'd would have a hard time (possible, but not easy and not completely polished).

    Comment


    • #22
      I like Linux can take whatever it wants from BSD and BSD can't. However, I don't like proprietary systems can do the same. There's no ability to keep something unique under BSD license.

      Comment


      • #23
        Originally posted by Templar82 View Post
        I'm not trying to start a software licence war, but I assume all their code is closed source and these improvements wont make it back to FreeBSD?
        Author here.. Every optimization is either (a) in head now, or (b) has a review pending. The slides reference the FreeBSD SVN revision numbers where patches landed, and the differential review numbers where the open reviews are.

        Comment


        • #24
          Originally posted by mrugiero View Post
          It depends. In this case, they touch NUMA and the in-kernel network stack. On Linux both of them would clearly be derivative work, and covered by the GPL, so no secret source. They could circumvent this by using DPDK instead, but if they need to share the NICs between several processes (I'd expect them to be dedicated for those tasks) they'd would have a hard time (possible, but not easy and not completely polished).
          The kernel is GPLv2. They only have to share the source with people who legally received the corresponding binaries. That's the whole reason the AGPL was created.

          What you're thinking of does exist though... it's the reason that Open Watcom C/C++ can't get into Debian or Fedora package repositories. (The Sybase Open Watcom Public License requires you to publish the source to your changes to the compiler, even if you're only distributing the modified binaries within your company. Debian and Red Had have decided that doesn't meet their definition of a Free Software License... I assume because it is a de facto violation of the "No discrimination against fields of endeavor, like commercial use." clause in the Debian Free Software Guidelines.)

          Comment


          • #25
            Originally posted by kd__ View Post
            Anyone understand why there are different performance numbers in the abstract than in the presentation? I can't really find any reference of those in the slides which tell a different story than the abstract below.

            From the abstract:
            "I will present data from real Netflix servers showing an improvement of almost 2x on AMD EPYC (85Gbs -> 165Gbs), and 1.3x on Intel Xeon (140Gb/s -> 180Gbs). "

            Author here: That's something I have wanted to figure out. The prototype "evolved" during the development from a Naples to a Rome CPU. After investigating several potential software causes for the lower (68 vs 85) number, I came to the conclusion that Rome was just faster, which gave it more opportunity to contend for fabric bandwidth and kernel locks.

            Comment


            • #26
              Originally posted by kd__ View Post
              Anyone understand why there are different performance numbers in the abstract than in the presentation? I can't really find any reference of those in the slides which tell a different story than the abstract below.

              From the abstract:
              "I will present data from real Netflix servers showing an improvement of almost 2x on AMD EPYC (85Gbs -> 165Gbs), and 1.3x on Intel Xeon (140Gb/s -> 180Gbs). "

              In the video he mentions that they've upgraded the CPUs while working on the software side too. So you have hard- and software changes that affected the numbers. Their software patches for example dropped Xeon from 140 to 100 when not using their setup/configuration.

              Comment


              • #27
                Anyone know what the kTLS mentioned in the slides is about, and where can I find documentation on how to use it? Or at least some code samples?

                Comment


                • #28
                  Originally posted by xception View Post
                  Anyone know what the kTLS mentioned in the slides is about, and where can I find documentation on how to use it? Or at least some code samples?
                  Kernel TLS slides are video's talk.


                  Comment


                  • #29
                    Originally posted by log0 View Post

                    In the video he mentions that they've upgraded the CPUs while working on the software side too. So you have hard- and software changes that affected the numbers. Their software patches for example dropped Xeon from 140 to 100 when not using their setup/configuration.
                    Yeah but what's the point of submitting an abstract with information that appears to be in conflict (?) with the presentation? I'm asking because I'm pretty sure I'm missing something here. (OK The most likely reason then is they submitted the abstract long before the updates and did not/could not update it)
                    Last edited by kd__; 05 November 2019, 10:28 AM.

                    Comment


                    • #30
                      Originally posted by ssokolow View Post

                      Install uMatrix. You can configure the "*" (default) policy to be whatever you currently do, but then disallow the cookies column and allow the "1st-party" row group.

                      (from lowest to highest precedence, it's "columns, row groups, rows, cells" within a sheet, and the sheets have precedence from * to .com to phoronix.com to www.phoronix.com)
                      ...on mobile?

                      When I see a YouTube video I always open the article in an incognito tab, but if by accident I open it in the normal ones, I must clear all my cookies and cache.

                      Comment

                      Working...
                      X