OBASE: Object-Aware Address-Space Design to Fix Memory Tiering

Despite widespread deployment of memory tiering hardware and OS mechanisms, datacenters still overprovision DRAM. The root cause is hotness fragmentation: memory allocators place objects by size rather than access pattern, causing hot and cold objects to interleave within the same pages. A single hot object marks its page as active, trapping cold data in expensive DRAM. Analysis of Google datacenter workloads reveals that up to 97% of bytes in active pages are cold and unreclaimable.

OBASE addresses this with address-space engineering: a compiler-runtime system for unmanaged languages that acts as an object-aware frontend for page-aware OS backends such as kswapd, TMO, TPP, and Memtis. It tracks memory accesses through lightweight pointer instrumentation and migrates objects at runtime using a concurrency-safe, lock-free protocol, dynamically reorganizing the address space so hot objects cluster into uniformly hot pages and cold objects into uniformly cold pages.

Evaluated across ten concurrent data structures, six backends, and production traces from Meta and Twitter, OBASE lets unmodified tiering backends operate far more effectively—improving page utilization by 2-4x and cutting memory footprint by up to 70%, at only 2-5% overhead. For engineers, this points to a way to reclaim significant DRAM costs without modifying existing OS-level tiering infrastructure.