Carbink
Carbink is a research far-memory runtime that protects remotely stored spans with erasure coding.
Mechanism: Applications use remotable pointers. Equal-sized spans leave local memory in coded batches, but return individually; remote compaction reclaims the resulting holes. [1]
Failure boundary: Memory-node loss triggers reconstruction. Compute-node loss instead releases its remote allocations; application recovery requires a separate mechanism such as checkpointing. Byzantine and partial-network failures are outside the model. [1]
Design versus prototype: The design assumes a replicated memory manager. The evaluated prototype used an unreplicated manager, static assignment and no planned-failure handling. The paper omits detailed concurrent-failure handling. [1]
Evaluation scope: Eight same-rack machines used 50 Gbps NICs. Compared coding schemes ran inside the Carbink runtime; the EC-Split baseline follows Hydra's approach, not an independently deployed Hydra stack. These are historical research results, not general performance guarantees. [1]
Hydra contrast: Hydra splits each page into k data fragments and r parity fragments. In failure-recovery mode, writes wait for all k+r fragments; reads can reconstruct from the first k of k+delta requested fragments. Its EC-only mode completes after k writes without the same resilience guarantee. Availability therefore depends on the completion rule, not just the presence of parity. [2]
Engineering interpretation: Ask which component may fail, which metadata survives, and when protection becomes effective. Remote redundancy alone does not establish durable application recovery.
Relations
- related
- Erasure coding, Failure domain, AIFM, Hydra remote memory
Citations
- Zhou et al. Carbink. OSDI 2022, pp. 55-71; sections 3.1, 3.4-3.5, 4-5.
- Lee et al. Hydra. FAST 2022, pp. 181-198; sections 4.1.1-4.1.2.
Sources & caveats
- Primary-source synthesis; no native Carbink or Hydra execution, RDMA experiment or benchmark replication was performed.
- The final paragraph is an engineering interpretation, not a measured reliability result.