Warp-Level Divergence in GPUs: Characterization, Impact, and Mitigation

Ping Xiang, Yi Yang, Huiyang Zhou
Dept. of Electrical and Computer Engineering, North Carolina State University, Raleigh, NC, USA
20th International Symposium on High Performance Computer Architecture (HPCA’14), 2014

   title={Warp-Level Divergence in GPUs: Characterization, Impact, and Mitigation},

   author={Xiang, Ping and Yang, Yi and Zhou, Huiyang},



Download Download (PDF)   View View   Source Source   



High throughput architectures rely on high thread-level parallelism (TLP) to hide execution latencies. In state-of-art graphics processing units (GPUs), threads are organized in a grid of thread blocks (TBs) and each TB contains tens to hundreds of threads. With a TB-level resource management scheme, all the resource required by a TB is allocated/released when it is dispatched to / finished in a streaming multiprocessor (SM). In this paper, we highlight that such TB-level resource management can severely affect the TLP that may be achieved in the hardware. First, different warps in a TB may finish at different times, which we refer to as "warp-level divergence". Due to TB-level resource management, the resources allocated to early finished warps are essentially wasted as they need to wait for the longest running warp in the same TB to finish. Second, TB-level management can lead to resource fragmentation. For example, the maximum number of threads to run on an SM in an NVIDIA GTX 480 GPU is 1536. For an application with a TB containing 1024 threads, only 1 TB can run on the SM even though it has sufficient resource for a few hundreds more threads. To overcome these inefficiencies, we propose to allocate and release resources at the warp level. Warps are dispatched to an SM as long as it has sufficient resource for a warp rather than a TB. Furthermore, whenever a warp is completed, its resource is released and can accommodate a new warp. This way, we effectively increase the number of active warps without actually increasing the size of critical resources. We present our lightweight architectural support for our proposed warp-level resource management. The experimental results show that our approach achieves up to 76.0% and an average of 16.0% performance gains and up to 21.7% and an average of 6.7% energy savings at minor hardware overhead.
VN:F [1.9.22_1171]
Rating: 0.0/5 (0 votes cast)

* * *

* * *

Follow us on Twitter

HGPU group

1543 peoples are following HGPU @twitter

Like us on Facebook

HGPU group

274 people like HGPU on Facebook

* * *

Free GPU computing nodes at hgpu.org

Registered users can now run their OpenCL application at hgpu.org. We provide 1 minute of computer time per each run on two nodes with two AMD and one nVidia graphics processing units, correspondingly. There are no restrictions on the number of starts.

The platforms are

Node 1
  • GPU device 0: nVidia GeForce GTX 560 Ti 2GB, 822MHz
  • GPU device 1: AMD/ATI Radeon HD 6970 2GB, 880MHz
  • CPU: AMD Phenom II X6 @ 2.8GHz 1055T
  • RAM: 12GB
  • OS: OpenSUSE 13.1
  • SDK: nVidia CUDA Toolkit 6.5.14, AMD APP SDK 3.0
Node 2
  • GPU device 0: AMD/ATI Radeon HD 7970 3GB, 1000MHz
  • GPU device 1: AMD/ATI Radeon HD 5870 2GB, 850MHz
  • CPU: Intel Core i7-2600 @ 3.4GHz
  • RAM: 16GB
  • OS: OpenSUSE 12.3
  • SDK: AMD APP SDK 3.0

Completed OpenCL project should be uploaded via User dashboard (see instructions and example there), compilation and execution terminal output logs will be provided to the user.

The information send to hgpu.org will be treated according to our Privacy Policy

HGPU group © 2010-2015 hgpu.org

All rights belong to the respective authors

Contact us: