access icon free Memory Request Priority Based Warp Scheduling for GPUs

High performance of GPGPU comes from its super massive multithreading, which makes it more and more widely used especially in the field of throughputoriented. Data locality is one of the important factors affecting the performance of GPGPU. Although GPGPU can exploit intra/inter-warp locality by itself in part, there is still large improvement space for that. In our work, we analyze the characteristics of different applications and propose memory request based warp scheduling to better exploit inter-warp spatial locality. This method can make some warps with good inter-warp locality run faster, which is beneficial to improve the whole performance. Our experimental results show that our proposed method can achieve 24.7% and 11.9% average performance improvement over LRR and MRPB respectively.

Inspec keywords: multi-threading; parallel processing; graphics processing units; performance evaluation; scheduling; cache storage

Other keywords: memory request priority; warp scheduling; warps; super massive multithreading; intra/inter-warp locality; 11.9% average performance improvement; good inter-warp locality; inter-warp spatial locality; data locality; high performance; GPGPU; different applications; improvement space; GPUs

Subjects: Microprocessor chips; Microprocessors and microcomputers; Semiconductor storage; Computer vision and image processing techniques

http://iet.metastore.ingenta.com/content/journals/10.1049/cje.2018.05.003
Loading

Related content

content/journals/10.1049/cje.2018.05.003
pub_keyword,iet_inspecKeyword,pub_concept
6
6
Loading