From: ben@decadent.org.uk   
      
   On Fri, 2025-10-17 at 14:11 +0200, Gero wrote:   
   > Dear experts,   
   >    
   > thanks a lot for your work and commitment on the Debian system. I'm    
   > using Debian for years and I am generally very pleased with it.  :-)   
   >    
   > With my company we do numerical simulations and recently did some    
   > benchmarking tests on new AMD EPYC 9334 processors that showed a    
   > significant performance loss of a current Debian system compared to an    
   > older Red Hat or Rocky Linux. We could narrow that down to the following    
   > finding:   
   >    
   > Running these commands:   
   >    
   > cd /dev/shm   
   > echo 3 > /proc/sys/vm/drop_caches   
   > fio --name=random-write --ioengine=posixaio --rw=randwrite --bs=4k    
   > --numjobs=1 --size=4g --iodepth=1 --runtime=60 --time_based --end_fsync=1   
   [...]   
   > I would like to know if you have an explanation or an idea.   
      
   I think for this specific test, the explanation is "this is a stupid I/O   
   pattern that no-one optimises for".  Using AIO with a depth of 1 is   
   effectively doing synchronous I/O in a less efficient way.   
      
   Added to that, POSIX AIO was never that efficient on Linux, and the   
   upstream developers seem to have more-or-less given up on it in favour   
   of io_uring.   
      
   > And I wonder   
   > if you would be interested in investigating the issue any further. Or if   
   > you have a suggestion who I might address preferably.   
      
   If you can also see a regression for io_uring and a more sensible I/O   
   depth then this would probably be interesting for the upstream   
   developers,   
      
   Ben.   
      
   --    
   Ben Hutchings   
   Computers are not intelligent.	They only think they are.   
      
   -----BEGIN PGP SIGNATURE-----   
      
   iQIzBAABCgAdFiEErCspvTSmr92z9o8157/I7JWGEQkFAmj2cJEACgkQ57/I7JWG   
   EQkDQxAAxmggui5KCVKlkQK8zMg+FneICZOZWXqkNNUNFvgWkD0spl1HvqDw2Glj   
   EBqyF/t/MGgHJPhHarkRU2LDvrR7UyJdPtFpggNSXzQIj2XBBcCno1kPCq4VuEwS   
   QHxQ+7Ug3vjl5DLpjtucvWPyBSV9lXRp2T5KPzY6DUVjJDz6Xw4UmDLtHLE0osjt   
   1SrH93f3Nds1WToWXdkVmi7IfpJzHY77PNG0cfZ86OQ62ppHO2AyE+MaXIBGlMjY   
   jN5FbO0UTNtzjP7BDUHIBDh+YEv8kecHPbmMbNPfufxQlrUN9Egx8UUVw46Q11+2   
   Yt9nLBzsAG1vYNwB34B1sYZ9pM6uh/w8fA4yJwrwOE5YVWek9AJKAjyn+OKCZjHx   
   kRaJ9XnLXPaKZA3YANDe37omsVYAe6XsolvzxxhltHFcW0kYJCr3e2SFYDAgWVl6   
   EEDEfXm3nbsujlWbrcdn3DTCPpM3sRZ+uj2QPjRZw3vvmLCcAQMQDMf/TbrVPH3y   
   dQKzxFGm1hCCnCQrXlBAo99sol5rzwulOdDGuI592NE8xksJpzGfenLEHxeguTt4   
   gAcvz/LgYDvPyFdlNkwPJBwpIsX7VFfPgAOyzaWVPSDdJYEHYvPWJET52icmo+ki   
   +rXsXfiUlGTxVWpyur7sax631+oCe9CkXcikVEWF19+DGpMY2H8=   
   =OGYC   
   -----END PGP SIGNATURE-----   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)