Fastswap
Fastswap is a research Linux swap system that uses RDMA and modified page-fault handling to access remote memory without application changes.
Critical path: The faulted page is requested before speculative prefetches, with separate queues and polling for critical completions. Background reclaim reduces page-fault work but still needs CPU capacity. [1]
Paper versus code: Section 4.1 describes two queue pairs per CPU. The pinned driver instead allocates three: synchronous reads, asynchronous reads and synchronous writes. Queue-count claims therefore need a version, not just the system name. [1] [2]
Reclaim headroom: The paper permits a small excess above each cgroup's local-memory limit before direct reclaim resumes. A busy background reclaimer cannot make large allocations or abrupt limit reductions free. [1]
Swap-device trap: The prototype requires a registered swap device to activate kernel paths, but its documented setup sends no data there. That device is not evidence of a disk recovery copy. The code targets a patched Linux 4.11 kernel, not a current-kernel support contract. [3]
Performance scope: CFM combines Fastswap with a scheduler. Finishing a batch sooner can coexist with slower individual jobs; the nine-node comparison and forty-machine simulation are different evidence. [1]
Later diagnosis: DiLOS measures residual page-fault software costs even when reclaimed pages are clean. Excluding write-back traffic in that experiment does not eliminate allocation, accounting or reclamation work. [4]
Relations
Citations
- Amaro et al. Can Far Memory Improve Job Throughput? EuroSys 2020, article 14; sections 4, 5 and 6. DOI: 10.1145/3342195.3387522.
- Fastswap authors: RDMA driver queue types, CPU indexing and allocation, revision 9cfc2ad
- Fastswap authors: kernel requirements and swap-device configuration, revision 9cfc2ad
- Yoon et al. DiLOS: Do Not Trade Compatibility for Performance in Memory Disaggregation. EuroSys 2023; section 3.1, figure 1.
Sources & caveats
- Primary-source synthesis and pinned-code inspection, not a native Fastswap run, RDMA benchmark or failure-recovery experiment.
- The two-versus-three queue comparison is directly visible in the paper and source. It does not establish which code revision produced the paper's measurements.
- The absence of a disk recovery copy is an inference for the README's no-data-to-swap-device setup, not a claim about all remote-memory systems.