Announcement

Collapse
No announcement yet.

A Look At The Windows 10 vs. Linux Performance On AMD Threadripper 2990WX

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

  • #41
    Michael it would be interesting to see if there are any differences between home versions of Windows and Windows server due to tuning for latency vs throughput, better NUMA scheduling, or similar.

    Comment


    • #42
      Thank you Michael for the link you gave in your message, 08-15-2018, 12:50 AM.
      openbenchmarking.org showed two (2) interesting pages, comparing three different Linux kernels:
      Linux 4.18.0 & Linux 4.19 20180814, openbenchmarking.org/result/1808149-RA-LINUX419K08
      Linux 4.18 Git & Linux 4.17.14, openbenchmarking.org/result/1808126-RA-LINUX418K21

      Between these three Linux kernels are startling & significant benchmark differences. These are very much worth exploring, imho. Generally the results indicate that the better-debugged-optimized kernel (4.17.14) is faster than the newly released kernel (4.19 20180814). This is as much as we should expect.

      To see if this pattern is repeated, the same systems should be tested, with the following Linux kernels:
      v4.14.62, v4.15.18, ...

      The official Ubuntu site has numerous Linux kernels, ready compiled, for instant installation into any Debian-based operating system. These kernels include most (all) released by "The Linux Foundation", seconds after the source code is released.


      If we are restricted to "amd64", this is available in two version versions: "GENERIC" and "LOW LATENCY". Are these significant in bench-tests & other performances?
      The only "twist" in these compilations of the Linux kernel is that the separate COMPILED version of these are also released by the Debian group as well. Does this pure-Debian compilation perform differently from the Ubuntu version?

      If these Linux kernels are similar in the test results, we can assume that it is not the kernel that make the differences, but the other parts of the operating system.

      Comment


      • #43
        Not that should matter when comparing operating systems, but the Threadripper 2990WX uses 8 channel memory while Michael only gave it 4 DIMMs. That means that the numbers are worse than they should be if someone wants to use them to see how their existing system compares.

        Also, that remark about it not mattering assumes that Michael gave each Threadripper die 1 memory channel each. If he gave 2 dual channel and 2 nothing, it might give one OS an unfair advantage over another in a comparison depending on how they handle NUMA. Not that I would expect Windows to do better than Linux, but avoiding bias is important when doing benchmarks.
        Last edited by ryao; 15 August 2018, 11:13 PM.

        Comment


        • #44
          Originally posted by Qaridarium

          yes it is same effect than back in history with the AMD-Bulldozer release the bulldozer worked well and fast on linux but on windows the bulldozer cpu architecture failed.

          now it is same again windows people claim the new 32core amd cpu is slow but the linux people know that it is very fast cpu but windows is broken by design and can not handle this NUMA architectur cpu.
          To be fair, 7zip is partially to blame here for lacking NUMA support (I checked). There are APIs for specifying from where memory should be allocated and doing proper CPU affinities for the NUMA topology on both platforms:

          Numa memory allocation policy may be specified as a per-task attribute, that is inherited by children tasks and processes, or as an attribute of a range of process virtual address space.

          https://docs.microsoft.com/en-us/win...d/numa-support
          https://docs.microsoft.com/en-us/win...om-a-numa-node

          I cannot see anyone caring enough to actually use them for compression software though. In any case, Linux ought to handle NUMA better given that there are developers at various HPC companies working on improving its abilities on NUMA systems. Windows has basically no market share there and likely just has NUMA support as a me too feature, so hearing that it did worse on a NUMA system is no surprise.

          Comment


          • #45
            Originally posted by ryao View Post
            Not that should matter when comparing operating systems, but the Threadripper 2990WX uses 8 channel memory while Michael only gave it 4 DIMMs. That means that the numbers are worse than they should be if someone wants to use them to see how their existing system compares.

            Also, that remark about it not mattering assumes that Michael gave each Threadripper die 1 memory channel each. If he gave 2 dual channel and 2 nothing, it might give one OS an unfair advantage over another in a comparison depending on how they handle NUMA. Not that I would expect Windows to do better than Linux, but avoiding bias is important when doing benchmarks.
            No, all Threadrippers have just four memory channels. EPYCs have eight.

            Comment


            • #46
              Originally posted by Tomin View Post

              No, all Threadrippers have just four memory channels. EPYCs have eight.
              You are right. Thanks for pointing that out.

              Comment


              • #47
                This made forbes:

                Comment


                • #48
                  Ok Linux beats Windows for most benchmarks, that is somewhat expected.
                  I would like to see some benchmarks that include disk I/O.
                  My guess is that results would look even worse for Windows

                  Comment


                  • #49
                    It's the final prove SAP was biased toward Windows (and slowlaris in the past). This windows shit doesn't scale!

                    Comment


                    • #50
                      Originally posted by ryao View Post
                      I cannot see anyone caring enough to actually use them for compression software though. In any case, Linux ought to handle NUMA better given that there are developers at various HPC companies working on improving its abilities on NUMA systems. Windows has basically no market share there and likely just has NUMA support as a me too feature, so hearing that it did worse on a NUMA system is no surprise.
                      I remember your bullshit from the past and it seems you're spreading even more bullshit now. You're saying there's no Windows market share in HPC, but you didn't use the same argument when comes to Linux vs BSD. What's BSD market share in HPC? None. Furthermore, Windows is promoted by SAP (which was always biased) and handles quite big systems. Are you saying they're not using NUMA? What we're seeing here Windows is broken by design and doesn't scale. Oh, and there's no BSD in SAP, so it makes your arguments even more funny. You make no sense at all.

                      Comment

                      Working...
                      X