The Reading System

You have papers. The problem is not acquiring more of them — it is converting them into two things: a related-work chapter that makes an argument, and a working knowledge you can defend under questioning.

The trap this document exists to prevent is reading as a form of productive-feeling procrastination. Reading has no natural stopping point, always feels responsible, and can consume your entire timeline. The cure is to make reading produce artifacts. A paper you read without writing a note is a paper you will read again.

How much to read

For an MSc thesis, roughly 30 to 50 papers cited, of which 10 to 15 are core papers you know cold — methodology, exact claims, what they did not evaluate, why your work differs. Those are the ones your committee will ask about, and being fuzzy on a paper you cited is a bad moment in a defense.

The rest are contextual: you understand the contribution and why it is cited, without knowing every detail.

Set a reading budget in weeks, not papers. Roughly: heavy reading through weeks 1–8, then taper. Freeze related-work reading at about week 20. After that, you skim new arrivals only to check that nobody has just published your thesis. Otherwise a related-work chapter never converges, because the literature does not stop.

Read in three passes

Keshav's three-pass method, which is standard in systems and worth following exactly.

First pass, five to ten minutes. Title, abstract, introduction, section headings, conclusions, and a glance at the figures. Then answer: what category is this, what is the claimed contribution, is it correct-looking, and is it relevant to me? Most papers stop here, and that is the point — the first pass is a filter, and using it without guilt is how you cover ground.

Second pass, about an hour. Read the body, ignore proofs and heavy detail, look carefully at the figures and the evaluation setup. You should end able to summarize the paper with supporting evidence to someone else. This is where you write the note.

Third pass, several hours. Only for your core papers. Reconstruct the work: what would you have to build to reproduce this, what assumptions are load-bearing, where would it break. You do this for the 10 to 15 papers your thesis argues against.

For systems papers specifically, read the evaluation section with active suspicion. It is where the useful information hides. Which workloads are conspicuously absent? Is the baseline tuned or is it a default? Is the overhead of the mechanism itself reported, or only its benefit? How many runs, and is variance shown? Is the comparison against the strongest prior work or a convenient older one? The questions you learn to ask here are exactly the questions your committee will ask you, so it pays to internalize them early.

One structured note per paper

Use templates/paper-note.md. The schema matters more than the prose, because a fixed schema makes notes comparable and comparable notes are what turn into a literature matrix.

Non-negotiable fields, and why each one earns its place:

The citation key matching your BibTeX entry, so notes and citations stay linked.

The claim, in one sentence, in your own words. Copying the abstract defeats the purpose; the act of compression is what produces understanding.

How they evaluated it — workloads, baselines, hardware, metrics. This field is the one that pays off most, because it feeds directly into your own methodology decisions and into the "prior work did not measure X" arguments that often become your gap.

What they did not do — unmeasured costs, untested workloads, stated assumptions. This field is where thesis gaps come from. Be specific.

Relation to my work — do I build on this, do I compare against it, does it threaten my novelty? Mark threats explicitly. A note that says "THREAT: this is close to my idea" is worth a great deal in month eight.

One quotable sentence with a page number, so you can cite precisely without reopening the PDF.

Write the note in the same session you read the paper. A note written a week later is a note about your memory of a paper.

The literature matrix

Once you have fifteen or so notes, build a table: papers as rows, dimensions of comparison as columns. Use templates/literature-matrix.md.

The columns are the interesting part, because choosing them is the intellectual work. Pick the three to five dimensions along which your subfield actually varies — not generic ones. For a learned-scheduling thesis they might be: what is learned, where the model sits, decision latency, whether overhead is reported, workloads evaluated, whether generalization to unseen workloads is tested.

Choose your columns so that the gap becomes visible as an empty region of the table. This is also, usually, the single most valuable figure or table in your related-work chapter, and it doubles as a defense slide.

Related work is an argument, not a list

The most common weak chapter in an MSc thesis is a sequence of paragraphs each summarizing one paper in citation order. It reads like a book report and it establishes nothing, because a list has no thesis.

Organize thematically instead. Group papers by approach, and for each group: characterize what the group has in common, cite its members compactly, then state the specific limitation the group shares that your work addresses. The paragraph ends by advancing your argument.

The structural difference is easy to hear. A list sounds like: "Zhang et al. proposed X. Kumar et al. proposed Y. Lee et al. proposed Z." An argument sounds like: "One line of work learns policies offline from workload traces [Zhang, Kumar, Lee]. This makes inference cheap enough for the kernel hot path, but all three assume the deployment workload resembles the training distribution, and none reports behavior under workload shift — the regime this thesis targets."

Same citations, entirely different chapter.

Two mechanical rules that help. Every subsection ends with a sentence connecting back to your work; if you cannot write that sentence, the subsection may not belong. And never cite a paper you have not at least first-passed — an uncited-but-listed paper, or worse a mischaracterized one, is a real credibility problem if a committee member happens to be its author.

On placement: related work traditionally sits as chapter three, before your design. Some theses put it after evaluation so comparisons can reference actual numbers. Early is the safer default and what most committees expect. Follow your department's convention if it has one.

Citation management

Set this up before you have 40 papers. Retrofitting is hours of tedious manual work.

Zotero with the Better BibTeX extension is the standard choice. Configure it to auto-export your library to thesis/refs.bib on every change, and set a citation-key format you can guess without looking, such as authorYearShorttitle — zhang2023learned. Being able to type a citation key from memory is a real speed difference over a year.

Then fix your entries once, early. Auto-imported BibTeX is frequently wrong in ways that look sloppy in a bibliography: venue names inconsistently abbreviated, missing pages or DOIs, arXiv preprints cited when a peer-reviewed version exists, and capitalization mangled because BibTeX lowercases unless you protect it (write {OS} and {LLM}, not OS and LLM). Prefer the published version over the preprint whenever one exists.

For staying current without drowning: a few Google Scholar alerts on your core papers and key authors, the arXiv cs.OS and cs.DC feeds skimmed weekly at most, and DBLP for the proceedings of the venues that matter to you. Keep this to a bounded weekly slot.

The reading-to-writing bridge

Draft your related-work chapter in weeks 8 to 12, while the papers are fresh.

This feels early. It is not. You have read most of what matters by then, and writing this chapter early has three payoffs: it forces you to articulate your gap, which sharpens your thesis statement; it produces 8 to 12 real pages of your thesis while you are otherwise blocked on infrastructure; and it means you are not reconstructing forty papers from notes in month nine, which is the version of this task that people describe as miserable.

You will revise it later. That is fine — revising an existing chapter is a completely different order of work from writing one from nothing.