mind · research · also MMU In-Network for Disaggregated architectures

MIND

MIND is a research memory-disaggregation system that places address translation, protection and cache-coherence management in a programmable network switch.

Placement: MIND moves allocation policy to a switch CPU and address translation, protection and coherence metadata to its programmable data plane. Compute blades retain local DRAM; misses trigger page faults and RDMA requests. It is a shared-memory research system, not an SSD interface. [2]

Coherence trade-off: An MSI directory tracks regions that can contain multiple pages. A write can invalidate an unrelated page in the same region: false invalidation trades network work against directory capacity. [1]

Bounded splitting: A region's false-invalidation count above a threshold within an epoch splits a region in half, down to 4 KB pages. The analysis assumes stable access patterns and allocated pages over O(log2 M) epochs, with M measured in pages. This is not an unconditional bound for arbitrary changing workloads. [1]

Evaluation scope: Five physical servers emulate memory and compute blades with 100 Gbit/s connections. Fastswap and adapted GAM are comparison systems. Read-write contention and finite switch directory capacity limit scaling; the design targets one rack and one switch. [1]

Failure boundary: The paper discusses switch-state recovery and retransmitted invalidations, but does not introduce CPU/memory-blade fault tolerance. Inference: cache coherence is not a promise of crash-durable storage. [1]

Artifact limits: The pinned instructions require suitable programmable-switch and RDMA hardware, an NDA for switch programs and agreement to the RDMA driver's license terms. Trace-based experiments and PSO/PSO+ estimates are separate evidence; shortened traces need not reproduce the paper's exact values. [3]

Relations

related
Fastswap, Infiniswap, LegoOS

Citations

  1. Lee et al. MIND. SOSP 2021, pp. 488-504; sections 4.3-5, 7-8. DOI: 10.1145/3477132.3483561. (accessed 2026-11-15)
  2. MIND authors: control-plane, data-plane and local-DRAM architecture (accessed 2026-11-15)
  3. MIND authors: artifact evaluation, hardware, licensing prerequisites and simulation scope; pinned revision (accessed 2026-11-15)
  4. MIND authors: tested switch software dependencies; pinned revision (accessed 2026-11-15)
Sources & caveats
  • Primary-source synthesis and pinned artifact inspection, not a native MIND run, switch benchmark or failure-recovery experiment.
  • Coherence and durability answer different questions. The storage-durability distinction is an inference, not a reported MIND crash test.
  • The pinned guide tested Barefoot SDE 9.2 with grpc and thrift; it requires an Intel NDA and Mellanox OFED terms. These are historical prerequisites, not a current support contract. [4] [3]

JSON: /data/terms/mind.json

Published by Hesela team · Updated

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. MIND. Hesela. 2026-11-15. https://hesela.dev/terms/mind/