Announcement

Collapse
No announcement yet.

Latest Linux Code Smashes 14M IOPS Per-Core With Intel Core i9 12900K + Optane

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

  • #11
    Originally posted by sdack View Post
    Not everyone is a Putin. There are still good people in this world. Many volunteers in open source do not program for the money, but simply because they can and because it serves the greater good. Contributing code to make open source better is a wonderful thing in this crazy world. You should not have mixed feelings about it, but only good ones.
    Of course. As I said above, it is a good thing. I don't do a lot of IT things for the money either. Love them or hate them, I've volunteered for Mozilla for 12+ years. Again, not for the money but to help make the product better. Same goal.

    Comment


    • #12
      Important work, as this might facilitate the merge between persistant (SSD) and voltile (DDR RAM) storage.
      Sooner or later i believe technologies like 3DXpoint will reach performance levels where they can replace traditional RAM.
      That way we can have almost 0 seconds software loading times.

      First step could be to use mmap() files / block devices instead of malloc() physical ram for large objects (instead of swapping by the kernel).
      Second step we could look at a platform / CPU which allows mapping part´s / whole "block devices" into physical ram on CPU level / via Hardware components.
      When launching a process, we could instruct the kernel to only use "persistant" physical memory pages..
      There would be a function to "hybernate" the process, which would just remove all threads from the kernel + remove the virtual memory pages from the MMU.
      You can "unhybernate" the process later on (even after a reboot) by just re-mapping the memmory pages and re-adding the threads, without any "load" time.

      Comment


      • #13
        The bench runs at 512 byte block size right? Is there more than one disk in use or just the single one mentioned? The Intel product page says 1.5M IOPS at 4k block size and 7.2GB/sec read bandwidth, so for 512 bytes that's about 14M IOPS?
        Code:
        7.2e+09 / 512 =  14 062 500
        If it's just a single disk being tested currently, then he's already bottlenecked at the disk/interface not CPU core?

        ---

        Off-topic: The ads in these forum threads is nuts. Sometimes they get in the way of the UI buttons such as pagination, I just saw several static ones bunched up together 3x2 grid above this input box. The bottom right video ad somehow played with audio enabled without interaction from me which surprised me, it's X button was large but I could not close it by clicking on it, nor disable audio from clicking the icon. Had to mute browser/OS audio.

        Ironically though, my main motivation to block ads will be because of the heavy CPU usage. This is an old system but fully hitting one core for a sustained period is revving up the fan. I'm not sure if it's the animated ads or multiple video ones at the same time. I guess video is using software decoding or something.

        Comment


        • #14
          Originally posted by polarathene View Post
          The bench runs at 512 byte block size right? Is there more than one disk in use or just the single one mentioned? The Intel product page says 1.5M IOPS at 4k block size and 7.2GB/sec read bandwidth, so for 512 bytes that's about 14M IOPS?
          Code:
          7.2e+09 / 512 = 14 062 500
          If it's just a single disk being tested currently, then he's already bottlenecked at the disk/interface not CPU core?
          It is tested using multiple Intel Optanes

          Comment


          • #15
            Originally posted by fransdb View Post
            If indeed the same hardware has been used, it's indeed impressive.
            It appears to have been the same P5800X disk since the Phoronix 2.58M IOPS article in 2020 (although back then it's only referred to as an Intel Optane Gen2 disk), and last I recall it was an AMD ThreadRipper (nope, there was an article in Feb with this CPU, but last year he had switched to AMD), and some older Intel system prior to that. When CPU wasn't the bottleneck he added more of the same disk, otherwise if it was CPU he'd upgrade to a more powerful CPU.

            So I think he has 4 of these P5800X now? I thought earlier results were with a Optane 905P (costs about 25% the price for equivalent storage capacity and roughly a third of the IOPS/bandwidth) but was wrong. So is the 14M IOPS on a single core across all the disks at 512 block size?

            Comment


            • #16
              14 Mil Iops is like....

              mNA4dm.jpg
              Attached Files

              Comment


              • #17
                Originally posted by kozman View Post

                It's the Amazons and other hosting giants who will, IMHO, undeservedly reap the benefits of this by charging more, rather than less, for a benefit they didn't create.
                This is a good point, but let's not discourage Axboe from doing this important groundwork. In the end, we all benefit, whether for-profit or non-profit.

                Comment


                • #18
                  Originally posted by ermo View Post
                  GJ intel. Now, if you could just enable ECC support on the K-series...
                  Looks like i9-12900K on a W680 chipset supports ECC memory. Source.

                  Comment


                  • #19
                    It's a late post to this topic, but it's hot off the presses. Intel is closing Optane for good (https://www.servethehome.com/intel-o...022-wind-down/).
                    How long will we have to wait before another technology comes close to saturating a CPU with millions of transactions?

                    Or another thought: why wasn't Intel capable of turning this technology profitable? Arguably the technology is great. I love my Optane 905 SSDs. I have been aching to get my hand on a P5800X - but not at the current price points. Newegg is selling off 960GB 905s for about $600 a piece. The cheapest P5800X comes in at about $1200 for 400GB capacity.
                    Is the technology too advanced for the time? Most applications do well with NAND based SSDs, where an array of NAND SSDs is cheaper and has higher capacity than Optane at similar performance but arguably higher complexity (RAID) and lower density. Specialized applications (e.g. databases, telco level networking) benefit from Optane, but seemingly don't produce sufficient demand for the continued investment into this technology.

                    RIP Optane!

                    Comment

                    Working...
                    X