{
  "slug": "pond",
  "term": "Pond",
  "aliases": [
    "Pond memory pooling"
  ],
  "advantages": [],
  "limitations": [
    {
      "text": "The author artifact supplies emulation scripts, not the full Pond system. Its README warns against using memeater for memory-tiering tests because its pages can migrate.",
      "citationIds": [
        "artifact"
      ]
    }
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-16",
  "definition": "Pond is a research CXL memory-pooling design that assigns local and pooled DRAM to virtual machines using predicted latency sensitivity and untouched memory.",
  "note": "Pond preallocates a VM's memory rather than oversubscribing it. A memory-only virtual NUMA node (zNUMA) steers initial guest allocations toward local DRAM. Predictions choose the split; a QoS monitor mitigates mistakes. Each pool slice belongs to at most one host at a time. [paper]\n\nEvidence: Experiments cover 158 workloads using NUMA-based CXL emulation. Simulations use 75 days of requests from 100 clusters in 34 datacenters. Figure 21 reports 7% lower aggregate DRAM requirements for a 16-socket pool in the higher-latency scenario, at a 5% performance-degradation margin and 98% target tail percentage. This is a trace-driven simulation result, not a measured fleet saving. [paper]\n\nLatency denominator: local/remote pairs are 78/142 ns and 115/255 ns. Remote latency is about 1.82x and 2.22x local latency, or 82% and 122% higher. The paper's 182%/222% labels describe ratios, not additional increases of those amounts. They are not application slowdown. [paper] [examples]\n\nPooling scope: CXL 2.0 excludes mapping one physical memory location to multiple hosts simultaneously. Logical devices can share bandwidth without sharing the same bytes. This is version-specific, not a restriction on every later CXL version. [cxl2]\n\nCapacity check: aggregate free bytes do not ensure VM admission. Constructed examples separate free cores, local GiB, accessible pool GiB and an assumed minimum-local requirement; they do not predict that requirement from performance. [examples] [code]",
  "status": "research",
  "related": [
    "cxl-memory",
    "memory-stranding",
    "cxl-memory-pooling"
  ],
  "citations": [
    {
      "id": "paper",
      "title": "Li et al. Pond: CXL-Based Memory Pooling Systems for Cloud Platforms. ASPLOS 2023, pp. 574-587. DOI 10.1145/3575693.3578835; sections 2-6.",
      "url": "https://www.microsoft.com/en-us/research/wp-content/uploads/2022/10/2023_Pond_asplos23_official_asplos_version.pdf",
      "accessed": "2026-11-16"
    },
    {
      "id": "artifact",
      "title": "Pond author artifact README, commit 5ac0e66b3624b9324cc9042abcd930127005d4bc: emulation scope and memory-tiering warning",
      "url": "https://github.com/MoatLab/Pond/blob/5ac0e66b3624b9324cc9042abcd930127005d4bc/README.md",
      "accessed": "2026-11-16"
    },
    {
      "id": "cxl2",
      "title": "CXL Consortium: CXL 2.0 memory pooling Q&A, 21 May 2021; ownership and bandwidth",
      "url": "https://computeexpresslink.org/blog/cxl-2-0-specification-memory-pooling-questions-from-the-webinar-part-2-2391/",
      "accessed": "2026-11-16"
    },
    {
      "id": "examples",
      "title": "Hesela: constructed single-VM admission examples and latency-ratio arithmetic (JSON)",
      "url": "https://hesela.dev/research/pond-admission.json",
      "accessed": "2026-11-16"
    },
    {
      "id": "code",
      "title": "Hesela: executable admission predicate and latency arithmetic (JavaScript)",
      "url": "https://hesela.dev/research/pond-admission.mjs",
      "accessed": "2026-11-16"
    }
  ],
  "provenance": [
    "Primary-source review of the ASPLOS 2023 paper (CC BY 4.0), version-pinned author artifact and CXL Consortium explanation. No native Pond or CXL workload experiment was run by Hesela.",
    "The worked admission predicate is an original constructed example, not the Pond scheduler, trace replay, performance model or cost forecast. It excludes contention, topology, fragmentation, failure domains and reconfiguration delays.",
    "The pinned scripts disable remote CPUs and change host settings, including swap and ASLR. They were inspected, not executed. Reproduction requires an isolated, appropriately authorized system. [artifact]"
  ]
}
