{
  "slug": "kangaroo",
  "term": "Kangaroo",
  "aliases": [
    "Kangaroo flash cache"
  ],
  "definition": "Kangaroo is a tiny-object flash cache that combines a small log-structured staging cache with a larger set-associative cache.",
  "note": "KLog batches objects for KSet to reduce set rewrites while limiting the DRAM index to the staging tier. KSet also retains Bloom filters and hit metadata in memory; no full-object index does not mean no DRAM cost. [paper]\n\nThreshold admission counts objects mapping to a set, not bytes. The pinned simulator rejects undersized groups and offers rejected objects to log readmission; warm-up controls when admission filtering runs. Its write-factor calculation weights the set component by the admitted-byte fraction before adding log writes. [threshold] [cache]\n\nConstructed drain: forty 100-byte objects occupy one padded 4,096-byte log write. With threshold 2, forty singleton groups cause no set writes and retain no objects after the drain; twenty pairs cause 81,920 set-write bytes and retain forty objects. Lower write traffic alone therefore does not establish better caching. This example has no hits, readmission, metadata or FTL copying. [example]",
  "status": "research",
  "advantages": [],
  "limitations": [
    {
      "text": "The published production tests report application write rates, not NAND-programmed bytes, and their different request streams prevent direct comparison between the two admission-policy experiments.",
      "citationIds": [
        "paper"
      ]
    }
  ],
  "related": [
    "ripq",
    "write-amplification"
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-19",
  "citations": [
    {
      "id": "paper",
      "title": "McAllister et al. Kangaroo: Caching Billions of Tiny Objects on Flash. SOSP 2021, pp. 243-262, sections 3-5. DOI: 10.1145/3477132.3483568.",
      "url": "https://www.cs.cmu.edu/~beckmann/publications/papers/2021.sosp.kangaroo.pdf",
      "accessed": "2026-11-19"
    },
    {
      "id": "threshold",
      "title": "Kangaroo simulator, commit 48397846: threshold admission and rejected-group handling.",
      "url": "https://github.com/saramcallister/Kangaroo/blob/48397846af0b838a0431977bc7a4a2e1493fe181/simulator/admission/threshold.hpp",
      "accessed": "2026-11-19"
    },
    {
      "id": "cache",
      "title": "Kangaroo simulator, commit 48397846: insertion, warm-up and calcFlashWriteAmp.",
      "url": "https://github.com/saramcallister/Kangaroo/blob/48397846af0b838a0431977bc7a4a2e1493fe181/simulator/caches/mem_log_sets_cache.cpp",
      "accessed": "2026-11-19"
    },
    {
      "id": "example",
      "title": "Hesela. Synthetic log-and-set drain: bytes, admission counts and explicit unavailable metrics.",
      "url": "https://hesela.dev/research/kangaroo/evidence.json",
      "accessed": "2026-11-19"
    },
    {
      "id": "model",
      "title": "Hesela. Executable byte-accounting example (CC0).",
      "url": "https://hesela.dev/research/kangaroo/model.mjs",
      "accessed": "2026-11-19"
    },
    {
      "id": "tests",
      "title": "Hesela. Exact-byte and input-validation tests (CC0).",
      "url": "https://hesela.dev/research/kangaroo/model.test.mjs",
      "accessed": "2026-11-19"
    },
    {
      "id": "article",
      "title": "Hesela. Cache admission can lower writes by dropping objects.",
      "url": "https://hesela.com/analyses/ssd-overprovisioning-denominators/#kangaroo-admission",
      "accessed": "2026-11-19"
    }
  ],
  "provenance": [
    "Primary paper and pinned simulator source reviewed; no original trace replay or on-flash benchmark was performed.",
    "The byte ledger is an original, single-drain illustration, not an implementation or replication of Kangaroo. Its fixed write units do not describe every SSD."
  ]
}
