Hermit
Hermit is a research Linux swap system that uses feedback-directed asynchronous work to reduce remote-memory page-fault costs without changing application code.
Feedback, not unlimited background work: Hermit adjusts reclamation using memory pressure and page turnaround, batches swap-out work, and overlaps reversible work with reads. Speculative duplicate fetches still resolve ownership at the final page-table update; losing requests release their pages. [1]
Denominator audit: Section 6.4.1 reports 13.8 to 10.2 microseconds but calls the reduction 35%. From those rounded endpoints, (13.8 - 10.2) / 13.8 is 26.1%; 13.8 / 10.2 - 1 is 35.3%, an inverse-latency gain. These are different metrics, not contradictory timings. The executable arithmetic does not reproduce the experiment. [1] [5] [6]
Overhead is not a share: The introduction's 14-microsecond total and 9-microsecond RDMA time imply a 5-microsecond residual: 55.6% relative to RDMA, but 35.7% of total latency. Do not mix these rounded inputs with the later 13.8-microsecond breakdown. [1] [5]
Evaluation boundary: The study used a CPU server and memory server on 100 GbE; Fastswap was ported to the same kernel. Its six workloads and local-memory ratios delimit the evidence. Less local memory can still make reclaim threads compete with application threads. [1]
Prototype boundary: The pinned tree targets Linux 5.14-rc5 and cgroup-v1. hermit_init initializes control flags to false; the README enables selected flags separately. The remoteswap README also offers a local-DRAM debugging backend, which cannot establish network performance. A fake swap file alone establishes no recovery guarantee. [2] [3] [4]
Relations
- related
- Fastswap, Infiniswap, Canvas
Citations
- Qiao et al. Hermit: Low-Latency, High-Throughput, and Transparent Remote Memory via Feedback-Directed Asynchrony. NSDI 2023; sections 1, 4-6, figure 10.
- Hermit authors: pinned kernel, cgroup-v1 and explicit activation recipe
- Hermit authors: control flags, debugfs exposure and hermit_init, pinned revision 10dfa5c
- Hermit authors: remoteswap RDMA setup, fake swap partition and local-DRAM debugging backend
- Hesela: explicit denominators calculated from rounded published latency endpoints
- Hesela: executable latency-ratio arithmetic and validation
Sources & caveats
- Source inspection and arithmetic audit, not a native Hermit run, RDMA benchmark, crash test or performance replication.
- Endpoint calculations inherit the paper's rounding; raw observations and sampling uncertainty were not reconstructed. A reciprocal latency ratio is not a measured application-throughput increase.
- Pinned source and documentation describe inspected artifacts, not proof of the exact revision or runtime configuration used for the paper's experiments.