Cache Fusion Traffic Patterns in Oracle RAC Under Heterogeneous Node Loads

Abstract

Cache Fusion efficiency is critical to the scalability and performance of Oracle Real Application
Clusters (RAC), where global cache coherency depends on rapid block transfers and synchronized
ownership arbitration across nodes. This study evaluates Cache Fusion traffic patterns under
controlled heterogeneous node load conditions, introducing selective CPU throttling, memory
pressure, and resource contention to isolate the effect of node performance asymmetry on cluster
coherency. Results show that a single degraded node significantly increases global cache transfer
latency, hot block ownership churn, and GCS/GES queue depth, even when it processes a smaller
share of total workload. Performance metrics confirm that RAC throughput decreases by up to 37%
under heavy throttling, whereas removal of the impaired node restores stability and improves
throughput despite reducing cluster size. These findings establish that RAC scalability depends more
on latency uniformity across nodes than on node count or compute volume. The study concludes that
real-world RAC optimization must prioritize equalized node responsiveness, interconnect
determinism, and proactive latency anomaly detection to prevent cascading coherency stalls in mixed
workload deployments.