PK-OMLP: An OMLP Based k-Exclusion Real-Time Locking Protocol for Multi-GPU Sharing under Partitioned Scheduling

Abstract

With rapid development of Graphics Processing Units (GPU) technologies, GPUs are strongly motivated to be adopted in many real-time applications. However, it is still a challenging work to efficiently integrate multiple GPUs into multicore/multiprocessor real-time systems, due to many real world constraints caused by the non-real-time closed-source GPU drivers. To avoid timing violations, k-exclusive locking protocols are developed to arbitrate exclusive access to each of the multiple GPUs. In this paper, a novel k-exclusion real-time locking protocol is proposed to handle multi-GPU sharing under partitioned fixed priority (P-FP) scheduling. The proposed protocol improves the prior work, the Clustered k-exclusion O(m) Locking Protocol (CK-OMLP) from two aspects: first, it allows multiple task on each CPU processor to make use of GPUs simultaneously, which improves the flexibility and increases GPU utilization in average cases, second, a suspension-aware analysis is presented to improve the schedulability, where task acquisition delays and GPU executions are modeled as self-suspensions. Schedulability experiments indicate that the proposed protocol outperforms the CK-OMLP in most considered scenarios.