TPP: Transparent Page Placement for CXL-Enabled Tiered Memory

With increasing memory demands for datacenter applications and the emergence of coherent interfaces like CXL that enable main memory expansion, we are about to observe a wide adoption of tiered-memory subsystems in hyperscalers. In such systems, main memory can constitute different memory technologies with varied performance characteristics. In this paper, we characterize the memory usage of a wide range of datacenter applications across the server fleet of a hyperscaler (Meta) to get insights into an application's memory access patterns and performance on a tiered memory system. Our characterizations show that datacenter applications can benefit from tiered memory systems as there exist opportunities for offloading colder pages to slower memory tiers. Without efficient memory management, however, such systems can significantly degrade performance. We propose a novel OS-level application-transparent page placement mechanism (TPP) for efficient memory management. TPP employs a lightweight mechanism to identify and place hot and cold pages to appropriate memory tiers. It enables page allocation to work independently from page reclamation logic that is, otherwise, tightly coupled in today's Linux kernel. As a result, the local memory tier has memory headroom for new allocations. At the same time, TPP can promptly promote performance-critical hot pages trapped in the slow memory tiers to the fast tier node. Both promotion and demotion mechanisms work transparently without any prior knowledge of an application's memory access behavior. We evaluate TPP with diverse workloads that consume significant portions of DRAM on Meta's server fleet and are sensitive to memory subsystem performance. TPP's efficient page placement improves Linux's performance by up to 18%. TPP outperforms NUMA balancing and AutoTiering, state-of-the-art solutions for tiered memory, by 10-17%