Huge Pages, Tiny Hotspots
Huge pages exchange placement granularity for translation reach. The trade was attractive when most bytes within a mapping shared similar access intensity and destination. Fewer translations reduce page walk traffic and relieve TLB pressure; contiguous allocation can also simplify the memory path. The inherited shortcut is to assume that the best unit for translation is also the best unit for migration. In a tiered system, those two units can have very different economics and need not remain coupled.
Imagine an illustrative 2 MiB page containing one repeatedly accessed 64 KiB region. Migrating the entire page moves 32 times the hot data and occupies fast tier capacity with mostly cold bytes. Moving only the hot region may require splitting the mapping, sacrificing translation reach and paying page table and invalidation costs. Neither choice is intrinsically correct. The question is whether the latency and bandwidth recovered during the region's useful lifetime outweigh the complete cost of changing its placement.
Let A be expected future hot accesses and ΔL their exposed latency saving. Let M be migration and mapping overhead, and T the additional translation cost caused by finer mappings over the same interval. Fine placement wins in this simplified serialized model when AΔL > M + T. Whole page movement needs a separate estimate that includes cold byte traffic and the opportunity cost of occupied fast memory. These expressions are screening tests, not cycle accurate predictors: memory level parallelism, overlapping migration, and queueing determine how much latency actually reaches execution.
The architectural change is to separate heat observation from translation policy. Collect access intensity below the current migration unit where practical, retain enough history to estimate stability, and expose whether a mapping can be split or later coalesced. A placement controller should compare whole page migration, finer relocation, replication where semantics permit, and leaving the data alone. Hardware and operating system interfaces must make remapping costs visible. Aggressive movement based on a short burst can otherwise spend bandwidth twice: once entering the fast tier and again when another transient hotspot displaces it.
Evaluate translation misses, useful bytes moved, migration queueing, and fast tier occupancy together. A lower TLB miss rate can coexist with worse service latency because the chosen mappings stranded the wrong data near the processor. Conversely, finer placement can lose when walks dominate or hot regions move too quickly. The useful metric is work completed under a fixed memory budget, including the traffic generated by policy. Is your page size selected for address translation, or silently deciding which cold bytes must accompany every hot one?