Clio
Clio is a research disaggregated-memory system that combines a client library with hardware memory nodes for address translation, protection and remote access.
Architecture: CBoard separates a hardware data path from software control on ARM. CLib manages client-side request ordering. The design is not ordinary one-sided RDMA on a memory server. [1]
Address trade-off: Hash-bucket overflow is avoided when allocating virtual addresses, moving retries to allocation rather than lookup. A requested fixed address is not always available. [1]
Retry boundary: Memory nodes retain recent write identifiers and atomic results. The described buffer covers two retries; exhaustion is reported to the application, not an unlimited exactly-once guarantee. [1]
Failure boundary: Request retransmission is not recovery from a failed memory node. The paper leaves built-in failure handling and replicated-write primitives for future work. [1]
Measurement boundary: Figure 8 uses 1 KB requests and reaches a 9.4 Gbps end-to-end testbed ceiling. Figure 9 exceeds 110 Gbps for large requests using on-board traffic generation that bypasses the 10 Gbps port. These are different measurement paths. Figure 6 labels ASIC latency as projected. [1]
Reproduction scope: The guide selects microbenchmarks, not every application or on-board experiment. Switched and direct-connected setups differ; the split cable adds latency. [2]
Build boundary: The guide targets Vivado/HLS 2019.1.3, Scala, ARM cross-compilation and host software. Other Vivado versions may need IP changes. [3]
API naming: The header declares legomem_alloc, legomem_read, legomem_write_sync and legomem_write_async, not the paper's ralloc/rread/rwrite names. Declarations alone do not prove implementation guarantees. [4]
Relations
- related
- Infiniswap, RDMA storage networking
Citations
- Guo et al. Clio: A Hardware-Software Co-Designed Disaggregated Memory System. ASPLOS 2022, pp. 417-433; sections 4.2, 4.5, 5, 7, 9.
- Clio authors: artifact evaluation scope, selected microbenchmarks and two testbed configurations, pinned revision 7908c3a
- Clio authors: hardware, ARM and host compilation prerequisites, pinned revision 7908c3a
- Clio authors: public host API declarations, pinned revision 7908c3a
Sources & caveats
- Primary-source synthesis of the final paper and pinned author documentation/header. No FPGA, RDMA, Clio runtime or benchmark was executed by Hesela.
- Throughput numbers are author-reported, not Hesela measurements. No confidence intervals or general production-speed advantage are inferred.
- The code revision is a reproducible inspection target, not proof that it produced every result in the paper.