AIFM
AIFM is a research runtime that uses application-aware pointers and containers to move objects between local and remote memory.
Mechanism: C++ interfaces expose object access patterns to a runtime using TCP-connected memory. Data structures can guide prefetching, eviction and remote computation instead of relying only on page faults. [1]
Pointer lifetime: A DerefScope fences evacuation while local data is in use. Accessing a raw pointer obtained by dereferencing outside that scope is undefined behavior; the remotable pointer can remain alive after its object moves. [1]
Prototype limits: The pinned implementation permits only one live DerefScope per thread. Remote-space garbage collection is not implemented; its exhaustion handler only prints a warning. Local-memory reclamation does exist. [2] [3]
Deployment scope: The paper's prototype connects each compute server to one memory server, without cross-client sharing or runtime resizing. Its SSD backend has block-size amplification and cannot run the remote computations supported by the memory backend. [1]
Performance interpretation: The evaluation uses two 25 Gbit/s machines and workload-specific local-memory budgets. Local-hit overhead, remote access and end-to-end throughput are different measurements; none establishes universal local-RAM speed. [1]
Later comparison: Eden's authors identify object-access instrumentation and application changes as AIFM trade-offs. Their hybrid approach uses hints at selected code locations and hardware page faults elsewhere. This comparison concerns integration and access costs, not crash durability. [4]
Relations
Citations
- Ruan, Schwarzkopf, Aguilera and Belay. AIFM. OSDI 2020, pp. 315-332; sections 4.2.2, 7 and 8.
- AIFM authors: README, supported platform and known limitations, pinned revision
- AIFM authors: FarMemManager remote allocation and unimplemented far-memory GC, pinned revision
- Yelam et al. A Developer-friendly approach to Application-integrated Far memory. USENIX ;login:, March 2025.
Sources & caveats
- Primary-source synthesis, not an AIFM execution, performance replication or crash-recovery experiment.
- The paper and pinned source describe a research prototype. Later comparisons use their own workloads and should not be pooled as one benchmark.