Our research group will present a paper at EuroSys 2025! This research addresses the growing demand for larger main memory capacity by proposing a novel proactive demotion scheme for tiered memory systems. Our approach, called PET, leverages application-level memory allocation patterns to efficiently manage data movement between fast DRAM and slower, more cost-effective memory (e.g., CXL-DRAM). PET introduces a new demotion unit, the PET-block, which aligns with application memory allocation behavior, leading to significant reductions in fast memory usage. Experimental results demonstrate that PET achieves a 39.8% reduction (up to 80.4%) in fast memory usage on average with minimal performance impact, and outperforms existing tiered memory management schemes in mitigating performance degradation under high memory pressure.
This work was co-authored with Dr. Yaebin Moon (formerly of my research group at Seoul National University), now at Samsung Electronics) and Professor Eojin Lee at Inha University. The road to publication was quite bumpy for this paper as we experienced three rejects and two major revisions before being accepted. However, this painful progress helped us better shape the paper through thoughtful comments by the reviewers, which we truly appreciate.
It is my first paper at EuroSys, which will take place in Rotterdam co-located with ASPLOS 2025. Our group has two papers accepted at ASPLOS 2025 and several of my students will go to the Netherlands this April. I will do my best to join the party and interact with the bright minds conducting systems research.
PET: Proactive Demotion for Efficient Tiered Memory Management
Wanju Doh, Yaebin Moon, Seoyoung Ko, Seunghwan Chung, Kwanhee Kyung, Eojin Lee, and Jung Ho Ahn
Tiered memory is a promising approach for increasing main-memory capacity at a lower cost by using DRAM as the upper tier (fast memory) and slower-but-cheap byte-addressable memory as the lower tier (slow memory). A proactive demotion, one of the ways to use tiered memory efficiently, demotes cold data to slow memory even when fast memory has sufficient free space. Prior works have utilized proactive demotion to reduce the high cost of main memory by reducing applications’ resident set size in fast memory. Further, proactive demotion helps mitigate severe performance degradation caused by fast memory shortages when there is a spike in demand for hot data. Still, we observe that leveraging memory access locality within the allocation units of applications enables larger fast-memory savings with lower system overhead.
We propose a new proactive demotion scheme, PET, which performs proactive demotion for efficient tiered memory management. PET proposes extending the unit of demotion and promotion from the OS page, adopted by prior works, to PET-block (P-block), which reflects the unit in which applications allocate memory. We also provide the mechanisms that carefully select the demotion target P-block and swiftly promote the demoted P-block when the access pattern changes. The prototype of PET on Linux kernel v6.1.44 reduces 39.8% (up to 80.4%) of fast-memory usage with only a 1.7% performance drop on average of the evaluated workloads. Also, it mitigates 31% performance slowdown compared to the default Linux kernel when the system’s memory usage is larger than fast-memory capacity, which outperforms state-of-the-art schemes for tiered memory management.