Locality-Based Cache Management and Warp Scheduling for Reducing Cache Contention in GPU.

Area of Science:

• High-performance computing
• Computer architecture
• Parallel processing

Background:

• Graphics Processing Units (GPUs) are crucial for high-performance computing but are limited by memory latency.
• Shared, small L1 data caches in GPUs lead to contention and pipeline stalls due to multiple warps.
• Inefficient cache management hinders overall GPU performance.

Purpose of the Study:

• To introduce Locality-Based Cache Management (LCM) and Locality-Based Warp Scheduling (LWS) to mitigate GPU cache contention.
• To enhance GPU performance by optimizing cache utilization and memory access patterns.
• To address the bottleneck of memory operation latency in GPU computing.

Main Methods:

• Classifying load instructions based on data locality: streaming, intra-warp, and inter-warp.
• Implementing LCM to manage cache behavior according to data locality.
• Utilizing LWS to apply cache bypass for streaming data and enhance inter-warp memory coalescing.

Main Results:

• LCM and LWS effectively reduce cache contention and pipeline stalls.
• Improved cache utilization through intelligent cache bypass and request coalescing.
• Achieved an average performance improvement of 26% compared to baseline GPU performance.

Conclusions:

• LCM combined with LWS significantly enhances GPU cache performance.
• The proposed methods offer a viable solution to memory latency bottlenecks in GPUs.
• This approach leads to substantial overall performance gains in high-performance computing applications.