{
  "slug": "amp-buffer-caching",
  "term": "AMP buffer caching",
  "aliases": [
    "Adaptive Multi-Policy caching",
    "Program-context-specific buffer caching"
  ],
  "definition": "AMP is a research buffer-caching framework that selects replacement policies and adapts cache partitions using the program contexts issuing I/O.",
  "note": "Requests are grouped by call-stack context, rather than only by file or process. Detected looping contexts use MRU partitions; the design discusses ARC as the default and adapts partition sizes using randomized competition. Repeated accesses alone do not distinguish looping from temporal clustering. [paper]\n\nThe reference-recency score uses the previous occurrence's position p in the oldest-to-newest list of distinct previously seen blocks: p/(n-1). First occurrences are excluded; a repeated block with n=1 receives 0.5. The reported loop threshold is 0.4, not a universal workload boundary. [report]\n\nHesela's synthetic examples [1,2,3,1,2,3] and [1,1,2,2,3,3] have identical lengths, per-block counts and three repeat observations. Their mean scores are 0 and 5/6 respectively. With no repeat observations the mean is undefined, not zero. These calculations distinguish order from frequency; they are not cache hit rates or measured speedups. [model]\n\nThe Linux 2.6.8.1 implementation retains a CLOCK-like default partition, so the design's ARC discussion must not be silently substituted for the kernel baseline. The evaluation's PCC* comparator combines PCC detection with AMP cache management, rather than reproducing the complete PCC system. [report]",
  "status": "research",
  "advantages": [],
  "limitations": [
    {
      "text": "The educational calculation omits context collection, sampling, partition competition and actual I/O; it does not replicate AMP's kernel prototype.",
      "citationIds": [
        "model",
        "report"
      ]
    }
  ],
  "related": [
    "adaptive-replacement-cache",
    "lru-replacement",
    "locality-of-reference",
    "program-counter-based-classification"
  ],
  "supersedes": [],
  "supersededBy": [],
  "firstSeen": null,
  "lastSeen": null,
  "updated": "2026-11-21",
  "citations": [
    {
      "id": "paper",
      "title": "Zhou, von Behren and Brewer. AMP: Program Context Specific Buffer Caching. USENIX ATC 2005, sections 1-3.",
      "url": "https://static.usenix.org/events/usenix05/tech/general/full_papers/short_papers/zhou/zhou_html/index.html",
      "accessed": "2026-11-21"
    },
    {
      "id": "report",
      "title": "Zhou, von Behren and Brewer. Program Context Specific Buffer Caching with AMP. UCB/CSD-05-1379, April 2005, sections 3.1, 3.3 and 4.2.",
      "url": "https://www2.eecs.berkeley.edu/Pubs/TechRpts/2005/Archive/CSD-05-1379.pdf",
      "accessed": "2026-11-21"
    },
    {
      "id": "model",
      "title": "Hesela. Bounded reference-recency calculation for synthetic integer block traces.",
      "url": "https://hesela.dev/research/amp-context/model.mjs",
      "accessed": "2026-11-21"
    },
    {
      "id": "tests",
      "title": "Hesela. Reference-recency boundary cases and independent last-occurrence oracle.",
      "url": "https://hesela.dev/research/amp-context/model.test.mjs",
      "accessed": "2026-11-21"
    }
  ],
  "provenance": [
    "Primary-source synthesis and a transparent arithmetic illustration, not a modern storage benchmark or independent replication of reported performance.",
    "Historical implementation and comparison qualifications are taken from the full technical report, not inferred from the abstract. [report]"
  ]
}
