Workload Scheduling in Heterogeneous Computing Environments

The patent segments the heterogeneous computing system into distinct processor types (CPUs, GPUs, FPGAs) and storage devices (HDDs, SSDs), creating separate task graphs and resource graphs that map specific computing tasks to appropriate processor types. This segmentation allows the scheduler to handle each processor type independently according to its characteristics, reducing overall scheduling complexity while maintaining performance benefits.

The patent applies local quality by considering data locality - storing data on storage devices physically proximate to the processors that need them. The scheduler evaluates storage device proximity to processors and assigns tasks to processor-storage pairs that minimize data transfer distances. This local optimization reduces data access latency and improves computing performance without requiring complex global coordination.

The patent performs preliminary actions by pre-positioning data on storage devices that are proximate to the processors that will need them. The scheduler analyzes task requirements and proactively assigns data to appropriate storage locations before execution begins. This preliminary data placement ensures fast data access during task execution without requiring complex real-time data movement operations.

The patent uses copying by creating multiple copies of data across different storage devices and processors. Instead of moving data dynamically, the system replicates data copies to storage devices proximate to various processors, allowing any processor to access its local copy without coordination overhead. This copying strategy improves data access speed while keeping the system relatively simple.