Overview
This series explains how Linux exposes CXL memory as a tier, why software sampling and fixed counters can miss the coverage, timeliness, granularity, flexibility, or overhead a policy needs, and how NEMO uses bounded controller-side aggregation for page, sub-page, and tenant views. Reported gains in HeMem, MEMTIS, and noisy-neighbor control are tied to named workloads and an FPGA prototype; migration, actuation, finite SRAM, and security boundaries remain.
Learning outcomes
- CXL makes placement consequential because local DRAM and attached memory can differ in latency and bandwidth; capacity alone is not enough.
- Useful telemetry must match a policy's required coverage, timeliness, granularity, flexibility, and overhead.
- NEMO lets the OS configure bounded controller-side views for huge pages, sub-pages, or tenants while the original data path continues.
- NEMO's reported gains depend on named workloads and an FPGA prototype; finite SRAM, revisit time, actuation, expressiveness, and ownership constrain transfer.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
Why does CXL make memory placement a telemetry problem?
CXL makes placement a telemetry problem because local DRAM and CXL-attached memory can differ in latency and bandwidth, so capacity alone cannot identify the right tier for changing hot pages or tenants.
How does NEMO differ from software sampling and fixed counters?
NEMO observes request metadata at the memory controller and lets the OS define bounded aggregations, while software methods spend CPU work and fixed counters often expose coarser, predetermined views.
What can NEMO measure for memory policies?
NEMO can maintain per-hugepage counts, base-page distributions inside selected huge pages, and many-pages-to-one-tenant counters; it does not provide arbitrary analytics.
Do NEMO's reported performance gains generalize to all CXL systems?
No universal result is established: the gains are reported for named workloads, baselines, configurations, and a CXL-attached FPGA prototype, with finite SRAM and software actuation still limiting outcomes.