The reading archive / 10 lessons
Morning
learning
Worked examples, experiments, and explanations to read at your own pace.
Why a GPU needs so many threads
Follow one matrix product through four resident warps, pending loads, and ready instructions; an interactive schedule separates latency hiding from faster memory.
GPU Architecture · Matrix multiply meets the hardware
Reservoir computing: let the system move, learn how to read it
Build reservoir computing from a stream of bits: fading memory, nonlinear features, a trained readout, and an experiment that removes one ingredient at a time.
Machine Learning / Physical Computing
BBR: how fast is the path, really?
Build BBR from one path: distinguish rate from flight volume, measure capacity through ACKs, probe deliberately, and confront the limits of an end-to-end model.
Computer Networks
How does 100 ms become a preemption?
Follow one ITIMER_VIRTUAL quantum through CPU-time accounting, signal delivery, saved registers, and a user-level scheduler—then expose the process-wide timer trap.
Operating Systems
What does ar actually build?
Trace real C files from relocatable objects into a static archive, then step through how symbol demand and command order decide which members reach the executable.
Systems Toolchain
How can an idle CPU be overloaded?
Follow two lock-protected paths to see why global utilization and delay signals conflict, then derive Protego’s split between throughput-driven admission and latency-aware lock queues.
Advanced Operating Systems / Research
Why Breakwater deliberately overcommits
Follow one credit ledger from safe underutilization to deliberate overcommitment, then see how queue-delay feedback and AQM contain the resulting incast risk.
Distributed Systems
Why delay grows like 1 / (capacity − load)
Derive the reciprocal delay curve from accumulated work, a geometric backlog, and Little’s law—then keep utilization fixed and remove the randomness.
Computer Networks
Why can a correct spinlock scale badly?
Trace how a lock cache line becomes a coherence hotspot, why test-test-and-set only partly helps, and how MCS queue locks turn global contention into directed local handoff.
Advanced Operating Systems
What converges in AIMD?
Follow two senders from unequal windows to fair sharing: a worked feedback cycle, linked plots, and the distinction between convergence and continuing motion.
Computer Networks