Skip to content
Snippets Groups Projects
Select Git revision
  • master default protected
  • release/202005
  • release/202001
  • release/201912
  • release/201911
  • release/releaseWindowsTestOne
  • release/windowsReleaseTest
  • release/releaseTest
  • release/releaseWindowsTest
  • release/201910
  • release/qt/201910
  • release/windows-test/201910
  • release/201908
  • release/201906
  • release/201905
  • release/201904
  • release/201903
  • release/201902
  • release/201901
  • release/201812
  • 4.0.0
  • 2.2.0
  • 2.1.0
  • 2.0.1
  • 2.0.0
  • 1.4.1
  • 1.4.0
  • 1.3.0
  • 1.2.0
  • 1.1.0
30 results

media_decoder.cpp

Blame
    • Adrien Béraud's avatar
      c2c6ddc0
      decoder: limit number of threads · c2c6ddc0
      Adrien Béraud authored and Guillaume Roguez's avatar Guillaume Roguez committed
      
      Using too many threads for decoding is not recommended and can
      introduce latency.
      Also during a call there are 2 live encoding and 2 live decoding,
      so using all threads for every decoding is not useful and could
      decrease performance and increase latency because of more
      CPU context switch.
      This is especially visible on machines with many logical cores
      like AMD Threadrippers or some Intel CPUs.
      Ideally, loads should use in total as many threads as logical cores
      to profit from the full machine power while limiting context switches.
      
      In this patch, decoders will use half the number of logical core,
      with a maximum of 8 threads.
      
      Change-Id: Ibe6c083c35fba972d930346629de2175625afd8c
      Reviewed-by: default avatarGuillaume Roguez <guillaume.roguez@savoirfairelinux.com>
      c2c6ddc0
      History
      decoder: limit number of threads
      Adrien Béraud authored and Guillaume Roguez's avatar Guillaume Roguez committed
      
      Using too many threads for decoding is not recommended and can
      introduce latency.
      Also during a call there are 2 live encoding and 2 live decoding,
      so using all threads for every decoding is not useful and could
      decrease performance and increase latency because of more
      CPU context switch.
      This is especially visible on machines with many logical cores
      like AMD Threadrippers or some Intel CPUs.
      Ideally, loads should use in total as many threads as logical cores
      to profit from the full machine power while limiting context switches.
      
      In this patch, decoders will use half the number of logical core,
      with a maximum of 8 threads.
      
      Change-Id: Ibe6c083c35fba972d930346629de2175625afd8c
      Reviewed-by: default avatarGuillaume Roguez <guillaume.roguez@savoirfairelinux.com>
    Code owners
    Assign users and groups as approvers for specific file changes. Learn more.