Cache-Only Memory Architecture¶
Make distributed node-local main memories cache-like, home-free stores that migrate and replicate shared data according to demand.
Core Idea¶
A cache-only memory architecture (COMA) gives a multiprocessor a shared address space while making each node's large local main memory act as an attraction memory cache. Data have no fixed physical home. Reads can draw or replicate data near the processors using it; the system must locate copies, maintain coherence, and avoid evicting the last remaining copy.[ref-2d995e175afe][ref-8711405bef23]
Scope of Application¶
COMA targets distributed shared-memory machines where remote accesses are costly. It is a reusable design class, not just the Data Diffusion Machine (DDM). DDM used hierarchical location machinery; KSR-1's ALLCACHE used a ring-of-rings machine with variable initial access time while a requested datum was found and brought local. Neither example proves COMA always outperforms other memory designs.[ref-2d995e175afe][ref-ca4d04e7ef13]
Clarity¶
An address identifies data logically, not a permanent home node. COMA differs from fixed-home NUMA, even if that NUMA machine also has processor caches. It also differs from an ordinary small CPU cache, whose evicted lines remain in separate backing memory. COMA's distributed attraction memories collectively are the main store.[^ref-2d995e175afe]
Manages Complexity¶
Automatic migration can make later accesses local without explicit placement by a programmer. It shifts work into directory/location tracking, coherence, tags, replication capacity, and safe replacement. More copies help readers but use memory, and a first remote miss still pays search and transfer costs.[ref-8711405bef23][ref-ca4d04e7ef13]
Abstract Reasoning¶
When a processor misses on a shared block, a location protocol finds a valid copy and moves or copies it into that processor's attraction memory. Coherence then coordinates later readers and writers. Before replacing a block, the system checks whether it would discard the only copy; if so, it preserves the data elsewhere. The location topology may vary, but home-free placement and last-copy safety cannot be omitted.[^ref-2d995e175afe]
Knowledge Transfer¶
DDM and KSR-1 use different interconnects but share this architecture-level pattern. Their performance policies do not automatically transfer. A content cache, demand-paged system, or ordinary shared-memory API is an analogy rather than COMA unless the distributed main memory itself is cache-organized and home-free.
[^ref-2d995e175afe]: Erik Hagersten, Anders Landin and Seif Haridi, “DDM—A Cache-Only Memory Architecture”, IEEE Computer 25(9), 44–54 (1992), abstract and design.
[^ref-ca4d04e7ef13]: Emilia Rosti, Evgenia Smirni, T. D. Wagner, Amy W. Apon and L. W. Dowdy, “The KSR1: Experimentation and Modeling of Poststore”, Oak Ridge National Laboratory report ORNL/TM-12287 (February 1993), abstract and architecture.
[^ref-8711405bef23]: Truman Joe and John L. Hennessy, “Evaluating the Memory Overhead Required for COMA Architectures”, Proceedings of ISCA (1994), abstract and introduction.
Neighborhood in Abstraction Space¶
Cache-Only Memory Architecture sits in a moderately populated region (56th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Distributed Systems Theorems & Fallacies (19 abstractions)
Nearest neighbors
- Memory Management — 0.86
- Buffer Overflow — 0.86
- Oblivious RAM — 0.86
- Fragmentation (computing) — 0.85
- Array-access analysis — 0.85
Computed from structural-signature embeddings · 2026-10-08