canvas · research · also Canvas remote memory

Canvas

Canvas is a research remote-memory swap system that isolates application swap paths and adapts allocation, prefetching and RDMA scheduling to reduce interference.

Mechanism: Canvas separates swap partitions, caches and RDMA scheduling by application. Reserving swap entries reduces repeat allocation locking; pressure releases reservations. Kernel prefetching can call on a modified runtime for semantic patterns. Scheduling balances applications and the timeliness of demand versus speculative reads. [1]

Evaluation scope: NSDI 2023 experiments used two machines, 25%/50% local working-set budgets and ten runs per application. Isolation-only and fully optimized configurations are different treatments. These results do not establish durability or adversarial security isolation. [1]

Percentage-point audit: Figure 14(b), GraphX-CC, reports contribution 52.1% to 62.8% and accuracy 46.9% to 52.4%. These are +10.7 and +5.5 percentage points, or approximately +20.5% and +11.7% relative changes. The paper's nearby percentage wording should not merge these denominators. Neither metric is application speedup. Contribution divides swap-cache-hit faults by all faults; accuracy divides those hits by prefetches. [1] [6] [7]

Activation boundary: The pinned scheduler initialization waits for configuration. The configuration helper calls the custom policy and application-configuration syscalls; loading the module alone is not evidence that the scheduler is active. This is static source inspection, not an observed running kernel. [4] [5]

Reproduction boundary: The author instructions target Linux 5.5, cgroup-v1 and a specific MLNX_OFED release. The remote-swap instructions also offer BACKEND=DRAM for local-memory debugging; that path cannot measure network performance. Historical kernel-install and reboot recipes are not safe production-host setup advice. [2] [3]

Relations

related
Fastswap, Hermit

Citations

  1. Wang et al. Canvas: Isolated and Adaptive Swapping for Multi-Applications on Remote Memory. NSDI 2023, pp. 161-179; sections 4-6 and figure 14(b). (accessed 2026-11-16)
  2. Canvas authors: pinned supported environment and cgroup-v1 setup (accessed 2026-11-16)
  3. Canvas authors: scheduler activation and local-DRAM debugging backend (accessed 2026-11-16)
  4. Canvas authors: rswap_scheduler_init and deferred configuration (accessed 2026-11-16)
  5. Canvas authors: configuration helper and custom syscalls (accessed 2026-11-16)
  6. Hesela: explicit denominators from rounded figure 14(b) inputs (accessed 2026-11-16)
  7. Hesela: executable percentage-point and relative-change audit (accessed 2026-11-16)
Sources & caveats
  • Primary-source synthesis, static code inspection and deterministic arithmetic, not a native Canvas run or benchmark replication.
  • Published rounded percentages do not recover event counts, run-level variation or confidence intervals; no statistical significance is inferred.
  • Pinned repository contents do not prove the exact configuration used for the published experiments.

JSON: /data/terms/canvas.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. Canvas. Hesela. 2026-11-16. https://hesela.dev/terms/canvas/