Research Report on Deadline Miss Rate
Budget / Salary$30–250
TypeFreelance project
LocationRemote
Posted2 hours ago
Code
MINForecast: eviction variant that reads sim.forecast instead of sim.actual.
Two tests in test_sim.py: (a) MINForecast == MIN when forecast = actual; (b) assets absent from the forecast are evicted first.
plot.py: sweep figure with capacity normalized to the live set.
reproduce.sh: workload generation -> sweep -> plot -> summary.md, one command.
requirements.txt, pinned Python version. Confirm seed determinism: two runs, identical summary.md.
EVALUATION.md
Model assumptions, each with file:line: non-preemptive transfers; full download before play; space reserved at transfer start; playing asset pinned; lookahead fixed at 1 h; MIN reads the actual schedule.
Adding a policy: choose_victims / candidates interface; meaning of stop_on_reject.
Swapping workloads: CSV schema; running with forecast != actual.
Gotchas: rejected on-demand requests leave the queue; why PrefetchNaive+LRU gets worse with more bandwidth; why all curves converge at the live set, not the corpus.
Tests: what brute-force MIN, capacity invariant, and infinite-cache each guarantee.
Test_Slide.pptx (one each: why it matters, where to change the code, rough effort)
Forecast noise sweep
Static playlist as the no-information forecast
Audience-driven workload with trigger classes
LRB-style learned next-use baseline
FOO-style bound
Skip-on-miss
Lookahead sweep
Demo (short, screen recording video)
Run reproduce.sh from a clean checkout.
Add a dummy policy; show it in the sweep next to the existing ones.
Walk through the three tests and what each one shows.
MINForecast: eviction variant that reads sim.forecast instead of sim.actual.
Two tests in test_sim.py: (a) MINForecast == MIN when forecast = actual; (b) assets absent from the forecast are evicted first.
plot.py: sweep figure with capacity normalized to the live set.
reproduce.sh: workload generation -> sweep -> plot -> summary.md, one command.
requirements.txt, pinned Python version. Confirm seed determinism: two runs, identical summary.md.
EVALUATION.md
Model assumptions, each with file:line: non-preemptive transfers; full download before play; space reserved at transfer start; playing asset pinned; lookahead fixed at 1 h; MIN reads the actual schedule.
Adding a policy: choose_victims / candidates interface; meaning of stop_on_reject.
Swapping workloads: CSV schema; running with forecast != actual.
Gotchas: rejected on-demand requests leave the queue; why PrefetchNaive+LRU gets worse with more bandwidth; why all curves converge at the live set, not the corpus.
Tests: what brute-force MIN, capacity invariant, and infinite-cache each guarantee.
Test_Slide.pptx (one each: why it matters, where to change the code, rough effort)
Forecast noise sweep
Static playlist as the no-information forecast
Audience-driven workload with trigger classes
LRB-style learned next-use baseline
FOO-style bound
Skip-on-miss
Lookahead sweep
Demo (short, screen recording video)
Run reproduce.sh from a clean checkout.
Add a dummy policy; show it in the sweep next to the existing ones.
Walk through the three tests and what each one shows.
Apply on Freelancer →
Project sourced from Freelancer.com. Applications happen directly on the original platform — we never collect your data.