GPU Preemption

• Preemption is the act of temporarily suspending a currently running task to execute a higher-priority task.
  
  
• In the context of GPUs, this means interrupting a GPU kernel or workload to switch to another one.
  
  
• This is typically managed by the GPU scheduler, which decides when and how to preempt tasks.
  

• Improved Responsiveness:
  Preemption allows high-priority tasks to be executed more quickly, reducing latency and improving the user experience.
• Resource Sharing:
  It enables efficient sharing of the GPU among multiple applications, preventing one application from monopolizing the resource.
• Fairness and Quality of Service:
  Preemption helps ensure that all applications get a fair share of GPU resources and that critical tasks are prioritized.

• Context Switching:
  Preemption involves saving the state of the current GPU task (its context) and loading the context of the new task.
  
  
• Timeslicing:
  Some preemption mechanisms use timeslicing, where each task gets a fixed amount of GPU time before being preempted.
  
  
• Wait-based vs. Reset-based:
  Some strategies wait for the current task to finish before switching, while others can interrupt and reset the task, requiring it to be idempotent (meaning it can be restarted without issues).
  
  
• Hardware and Software Support:
  Preemption requires both hardware support (e.g., mechanisms for context switching) and software support (e.g., GPU drivers and scheduling algorithms).
  

• Windows WDDM 1.2:
  Introduced a more robust GPU preemption model, ensuring that preemption requests are handled before a Timeout Detection and Recovery (TDR) process.
  
  
• Real-time Scheduling:
  Research focuses on enabling preemptive scheduling for real-time tasks on GPUs, allowing for predictable and timely execution of critical workloads.
  
  
• Multiprogrammed Workloads:
  Work is being done to extend GPU architectures to support efficient preemption in environments where multiple applications are running concurrently.
  

• Overhead: Preemption can introduce overhead due to the context switching process.
  
  
• Complexity: Implementing preemption can be complex, requiring careful design of scheduling algorithms and synchronization mechanisms.
  
  
• Idempotency: Reset-based preemption requires tasks to be idempotent, which may not always be the case.
  

abactuon said:

For some reason I thought it depended on

View attachment 141461

but I was probably wrong.

Click to expand...

shoober420 said:

Ive read conflicting results as to whether GPU Premption increases or decrease latency / FPS. i decided to make scripts to enable or disable GPU Preemption. There is this one script from AlchemyTweaks.

Officially-Verified/GPU Preemption/Disable Preemption.reg at main · AlchemyTweaks/Officially-Verified

Here you will find all tweaks that are backed by official vendor documentation (Microsoft, NVIDIA, Intel, AMD, etc.). Their behavior, purpose, and operational impact are formally specified, making ...

github.com

This only includes one registry key out of numerous others ive found pertaining to GPU Preemption. the script ive put together includes many more.

windows11-scripts/DisableGPUPreemption.bat at main · shoober420/windows11-scripts

Windows 11 Scripts. Contribute to shoober420/windows11-scripts development by creating an account on GitHub.

github.com

i also need to add other mentions of GPU Preemption from the main GPUTweaks script.

windows11-scripts/GPUTweaks.bat at main · shoober420/windows11-scripts

Windows 11 Scripts. Contribute to shoober420/windows11-scripts development by creating an account on GitHub.

github.com

my question is, does anyone know if these registry keys involve GPU Preemption and should they also be added to enable and disable GPU Preemption?

Code:

reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "ComputePreemption" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "DisableCudaContextPreemption" /t REG_DWORD /d "1" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableAsyncMidBufferPreemption" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableCEPreemption" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidBufferPreemption" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidBufferPreemptionForHighTdrTimeout" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidGfxPreemption" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidGfxPreemptionVGPU" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableSCGMidBufferPreemption" /t REG_DWORD /d "0" /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "PerfAnalyzeMidBufferPreemption" /t REG_DWORD /d "0" /f

they are under the "GraphicsDrivers" string, so they should naturally effect GPU Preemption, meaning it could possibly effect GPU Preemption. are those keys types of GPU Preemption? does anyone happen to know if those are needed to be enabled as well for GPU Preemption to work?

is there also a way to test if GPU Premption is on or off?

Click to expand...

Alchemy Tweaks said:

I tested

Code:

reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "ComputePreemption" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "DisableCudaContextPreemption" /t REG_DWORD /d "1" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableAsyncMidBufferPreemption" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableCEPreemption" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidBufferPreemption" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidBufferPreemptionForHighTdrTimeout" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidGfxPreemption" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableMidGfxPreemptionVGPU" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "EnableSCGMidBufferPreemption" /t REG_DWORD /d "0" /f<br>reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers\Power" /v "PerfAnalyzeMidBufferPreemption" /t REG_DWORD /d "0" /f`

From my perspective, I observed no measurable difference in FPS, latency, or related metrics. There may be internal GPU-level changes detectable through a deeper analysis (e.g., GPUView or similar tooling), but in my assessment, these do not deliver the performance improvements one might expect. The root cause is likely elsewhere. Υou may want to test ForcePreemptionAware. Based on my own experimentation with this set of settings ultimately keeping preemption enabled this DWORD showed some positive effect.

Click to expand...