Abstract

Emerging memory interconnect technologies, such as Compute 

Express Link (CXL), enable scalable memory expansion by inte- 

grating heterogeneous memory components like local DRAM and 

CXL-attached DRAM. These tiered memory systems offer potential 

benefits in bandwidth and capacity, but their heterogeneous per- 

formance characteristics pose significant challenges for efficient 

memory management. Traditional approaches, including hotness- 

based page placement, prioritize latency reduction by retaining 

frequently accessed data in near memory. However, these methods 

often fail to utilize aggregate memory bandwidth effectively, leading 

to imbalanced traffic and performance degradation under memory- 

heavy workloads. Additionally, the reliance on costly page migra- 

tions introduces overheads that impede responsiveness to workload 

variations. To address these challenges, we propose TierTune, a 

dynamic memory management framework integrating Last-Level 

Cache (LLC) partitioning and page migration. TierTune dynam- 

ically partitions the LLC between near and far memory nodes, 

rapidly mitigating latency imbalances and balancing memory traf- 

fic more effectively than software-based page migration. We extend 

this approach with a migration policy that dynamically redistributes 

pages to address persistent imbalances, incorporating cache hier- 

archy effects and bandwidth pressure into placement decisions. 

Experimental evaluations demonstrate that TierTune improves 

performance by 19.6% on average compared to state-of-the-art page 

migration policies across diverse workloads by balancing memory 

traffic more effectively and reducing migration overheads.