Companion archive — Distributed LLM Serving (CSUR submission)

This folder is the machine-readable companion to the survey "Distributed Serving Architectures for Large Language Model Inference: A Taxonomy, Quantitative Models, and Practitioner's Decision Framework" (Chukkapalli, Mishra, Naik — ACM Computing Surveys submission, 2026).

Deposit. Permanently archived on Zenodo under the concept DOI 10.5281/zenodo.20423002 (always resolves to the latest version; v1.0.1 specifically: 10.5281/zenodo.20423049). Source mirror at https://github.com/dchukkapalli-dev/distributed-llm-serving-companion.

It exists so that any reader can audit the taxonomy assignments and the PRISMA flow counts without re-deriving the source set, and can amend a pillar assignment by editing a single CSV row.

License (CC-BY-4.0)

Copyright (c) 2026 Divya Chukkapalli, Sanjay Mishra, Ganesh R Naik.

This companion is released under the Creative Commons Attribution 4.0 International License. You may share and adapt the contents for any purpose, including commercially, provided you cite the accompanying survey; see LICENSE for the full text and CITATION.cff for the required attribution format.

Contents

File	What it is
taxonomy.csv	All 56 included techniques with fields: name, pillar, representative_system, primary_objective, citation_key, cross_pillar. One row per technique. The pillar field is the chapter the technique is anchored in; cross_pillar lists secondary chapters where the technique is discussed but not anchored (so each technique is counted exactly once in totals). The citation_key matches an entry in references.bib.
references.bib	BibTeX records for all 120 cited works in the manuscript.
prisma_counts.csv	Stage-by-stage record counts for the PRISMA-2020-style flow diagram (Fig. 1 in the manuscript): identification (412), de-duplication (287), snowballing (+64), title/abstract screening (351 → 198), full-text eligibility (198 → 56).
CITATION.cff	Citation File Format (1.2.0) record so the artefact can be cited independently of the manuscript.
LICENSE	Full text of the CC-BY-4.0 license.

How to use

• Audit a pillar assignment. Open taxonomy.csv, locate the row for the technique, and check pillar and cross_pillar. The manuscript text and Section / chapter headings use the same pillar names.
• Recount included techniques. wc -l taxonomy.csv returns 57 (header + 56 rows); the 56 rows are exactly the techniques that passed full-text eligibility.
• Reproduce the reference count. grep -c '^@' references.bib returns 120.
• Reclassify a technique. Edit the relevant pillar / cross_pillar cell directly; downstream counts per pillar are a groupby on pillar.

How to reproduce the PRISMA flow

The stage-by-stage record counts in prisma_counts.csv and Fig. 1 of the manuscript were produced by the following procedure; a reader who repeats it on the same dates and venues should land within a handful of records of these numbers (small drift is expected as databases re-index).

1. Time window. 2019 through 2026; last refreshed May 2026. Pre-2019 foundational works (e.g. Vaswani 2017) are cited where structurally necessary but excluded from the PRISMA counts.

2. Venue list. Systems: OSDI, SOSP, NSDI, ATC, ASPLOS, ISCA, HPCA, SC, SoCC, MLSys. Machine learning: NeurIPS, ICML, ICLR, ACL, EMNLP. Preprints on arXiv were included only when their content was referenced verbatim by a later peer-reviewed paper or by a deployed engine.

3. Boolean query string.

   ("LLM serving" OR "language model serving" OR "inference serving" OR
    "transformer inference")
   AND
   ("distributed" OR "parallelism" OR "KV cache" OR "speculative" OR
    "disaggregation" OR "scheduling" OR "batching" OR "fault tolerance" OR
    "multi-tenant")
   

   Per-database adaptations: DBLP uses unquoted phrase syntax with & instead of AND; ACM Digital Library uses the Advanced Search AllField: qualifier; IEEE Xplore uses ("All Metadata":...); USENIX uses simple keyword search and is filtered by venue post-hoc; Google Scholar uses the same query verbatim with intitle: removed so that abstract matches are admitted.

4. Audit counts. Initial query: 412 records. After de-duplication across databases: 287 unique items. Forward + backward snowballing from the included set: +64 records (32 backward from references of included items; 32 forward via Google Scholar's citation index, last refreshed May 2026). Title/abstract screening: 351 → 198 retained. Full-text eligibility: 198 → 56 included as distinct techniques (the rows in taxonomy.csv). Background and adjacent-scope citations bring the total reference count in the manuscript to 120.

5. Snowballing stop criterion. Iteration stopped when one forward and one backward pass over the previously included set produced fewer than five new candidates that passed full-text eligibility---i.e. when the marginal yield of an additional iteration fell below 10% of the typical single-iteration yield.

Methodology note (single-coder)

Coding was performed by a single author. This release makes the coding trivially auditable; it does not replace a multi-coder inter-rater study, and the manuscript states this limitation in its Methodology section. The companion CSV's cross_pillar field is populated for five of the 56 techniques (~9%); these are the items most exposed to a second coder's disagreement and can be re-anchored by editing the corresponding pillar cell.

Licence (full text)

All files in this companion archive (taxonomy.csv, prisma_counts.csv, references.bib, CITATION.cff, this README.md) are released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0) to maximise reuse for follow-up surveys. The full license text is in LICENSE; the canonical short summary is at https://creativecommons.org/licenses/by/4.0/. Required attribution is a citation to the accompanying survey as specified in CITATION.cff.