Infiniswap
Infiniswap is a research remote-paging system that uses RDMA-accessible memory as swap space with local-disk fallback.
Design: A kernel block device serves paging requests while a user-space daemon manages donated RAM. Mapping uses slabs, but transfers operate on pages; no application rewrite is required. [2]
Read selection: A mapped slab is not proof that every page is remotely resident. The page bitmap is set after a remote write; reads otherwise use disk. Mapping alone does not prefetch old pages. [1]
Completion: A mapped page-out completes after RDMA WRITE, while its duplicate disk write is asynchronous. The shared dispatch buffer remains occupied until disk completion; exhausted entries can therefore block new requests. [1]
Failure scope: After remote loss, the design redirects reads to disk and uses the same slab's disk queue to order a pending write before its subsequent read. This is remote-loss handling, not a host-crash persistence guarantee. [1]
Prototype: The pinned README targets Linux 3.13.0 and reports tests on 4.4.0 and 4.11.0, requiring compatible RDMA support and an unused backup partition. These prerequisites are historical, not evidence of support on current kernels. [3]
Relations
- related
- Swapfile, RDMA storage networking, AIFM, Fastswap, Leap, Hermit, MIND, Clio, LegoOS, Hydra remote memory
Citations
- Gu et al. Efficient Memory Disaggregation with Infiniswap. NSDI 2017, pp. 649-667, sections 4.1, 4.3-4.5 and 8.
- Infiniswap authors: architecture and paging request paths
- Infiniswap author README, pinned prerequisites and component layout
Sources & caveats
- Source synthesis, not a native Infiniswap execution, crash test or benchmark replication.
- The host-crash distinction is an inference from the asynchronous disk path and the paper's remote-unreachability failure model. [1]
- Reported performance comes from specific 2017 RDMA workloads; section 8 assumes the network is not a bottleneck. [1]