Joint Dynamic Memory Allocation for Memory‐Intensive Applications in CXL‐Based Heterogeneous Memory Systems

ABSTRACT

The rapid growth of memory‐intensive applications poses significant challenges to memory capacity, bandwidth, and cost. Compute Express Link (CXL) offers a decoupled memory architecture and shows great potential for memory expansion and bandwidth enhancement in heterogeneous memory systems. We conducted experiments using typical memory‐intensive applications and observed that approximately 80% of memory‐intensive applications exhibit improved throughput with limited latency overhead. Throughput improvement depends on optimizing the DRAM‐CXL memory allocation ratio. To this end, we proposed Forced Memory Allocation and CXL Momentum‐Enhanced Progressive Adaptive Memory Optimization Algorithm (FA‐PAMO), which combines ratio‐guided memory allocation control and momentum‐enhanced adaptive optimization mechanisms. FA‐PAMO can quickly approach the optimal memory allocation ratio with fewer iterations compared to the dynamic allocation algorithm based on linear regression. Experimental results show that FA‐PAMO achieves a 33% faster convergence than the state‐of‐the‐art (SOTA) algorithm in dynamic memory allocation, reaching stability 3‐4 rounds earlier on average across seven representative workloads. It also improves throughput by approximately 29.14%.