From: nospam@needed.invalid   
      
   On Sun, 12/21/2025 1:28 PM, VanguardLH wrote:   
   > "Mr. Man-wai Chang"  wrote:   
   >   
   >> Announcing Native NVMe in Windows Server 2025: Ushering in a New Era of   
   >> Storage Performance | Microsoft Community Hub   
   >>    
   >>   
   >> ... more ....   
   >>   
   >> Modern NVMe devices—like PCIe Gen5 enterprise SSDs capable of 3.3   
   >> million IOPS, or HBAs delivering over 10 million IOPS on a single   
   >> disk—are pushing the boundaries of what storage can do. SCSI-based I/O   
   >> processing can’t keep up because it uses a single-queue model,   
   >> originally designed for rotational disks, where protocols like SATA   
   >> support just one queue with up to 32 commands. In contrast, NVMe was   
   >> designed from the ground up for flash storage and supports up to 64,000   
   >> queues, with each queue capable of handling up to 64,000 commands   
   >> simultaneously.   
   >>   
   >> .... more .....   
   >   
   > Wonder if it will reach non-Server editions, like Home or Pro.   
   > HKLM\SYSTEM\CurrentControlSet\Policies\Microsoft\FeatureManagement\Overrides   
   > doesn't exist on my Win11 Pro instance.  I doubt adding it would have   
   > any effect since it's very likely an option only for the Server edition.   
   > Speed is addictive, and users even of non-Server editions strive for   
   > best performance (that is bug-free and stable).   
   >   
      
   It's not ready quite yet. Patience.   
      
   For one thing, there is a drastic mismatch on queuing.   
   We will need to hear the story of what the plan is there.   
   Presumably the OS will query the drive, figure out from   
   its class that only a single queue of 32 is possible, and   
   then it will set up the interface that way for actual usage.   
      
   The purpose of queues at one time, was to allow the drive   
   CPU to compute the best order of completion for operations,   
   taking seek time into account. NVMe don't have conventional   
   seek time, and the access time is relatively constant   
   for each operation.   
      
   Where any whizzy interface comes into play, is if something   
   can be done in a "zero copy" way. That would buy some performance.   
   Whether zero copy is currently present, who knows.   
      
   But queuing by itself, I'm not convinced this buys a whole   
   hell of a lot for desktops. Maybe only a synthetic 4KB random   
   write test, benefits from the queue. ... and only for the   
   PCIe Rev5 NVMe with the 3 million IOPs rating. NTFS does   
   10,000 operations per second, on a good day, and deletes   
   in NTFS are extremely slow. I'll wait for the dog and pony show.   
      
      Paul   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)