async_cifar10: real-sim parity analysis, bug fixes, and parity checker - #63

Merged

dhruvsgarg merged 36 commits into

dg-fork-mainfrom

dg/streaming_expt_opt_gap

Jun 8, 2026

Commits on Jun 2, 2026

Streaming opportunity-gap experiment: mis-selection telemetry + oracle + harness

Dhruv Garg
and
claude
committed
Default experiment configs to time_mode=real

Dhruv Garg
and
claude
committed
Wire FedDance into async_cifar10 + fix base syncfl stack for fedavg/feddance

Dhruv Garg
and
claude
committed
Fix: REFL arms need aggr_num (OortSelector requires it)

Dhruv Garg
and
claude
committed
REFL: emit selection telemetry so the mis-selection oracle can score it

Dhruv Garg
and
claude
committed
Add FedDance arms (8-arm experiment) + 2-node, 2-phase run scripts

Dhruv Garg
and
claude
committed
Selector staleness audit: design doc + per-factor I_m staleness (step 1)

Dhruv Garg
and
claude
committed
Staleness audit (step 2, live): log believed scoring factors per candidate

Dhruv Garg
and
claude
committed
Staleness audit (step 3a): shared pure scoring fns + per-candidate score logging

Dhruv Garg
and
claude
committed
Staleness audit (step 3b): oracle counterfactual replay + true A_m

Dhruv Garg
and
claude
committed
Staleness audit (step 3c): self-relative regret in own-score units + cf reporting

Dhruv Garg
and
claude
committed
Per-run plot suite: PDF, config-stamped, grouped (performance/sanity/selection/insights/system)

Dhruv Garg
and
claude
committed
config_stamp: infer streaming from telemetry (snapshots omit config_overrides)

Dhruv Garg
and
claude
committed
Plot/telemetry refinements: labels, hist %, top config title, selection & runtime sanity

Dhruv Garg
and
claude
committed
Add max_runtime_s wall-clock cap (3.5h) to all experiment arms

Dhruv Garg
and
claude
committed

Commits on Jun 3, 2026

Fix streaming recv_fifo active-task leak that stalled async runs

Dhruv Garg
and
claude
committed
Update stale launch tests for baseline-owned aggregator_main

Dhruv Garg
and
claude
committed
Split node1 STREAMING batch into per-baseline YAMLs (felix, oort)

Dhruv Garg
and
claude
committed
Eliminate per-round trainer overhead that inflated "compute" time

Dhruv Garg
and
claude
committed
Task 3: real/sim parity hardening — seeding, determinism fix, checks, tests

Dhruv Garg
and
claude
committed

Commits on Jun 4, 2026

Sync sim-ordering for real/sim parity across sync baselines + doc

Dhruv Garg
and
claude
committed
random selector: work on sync FL too; + logical commit-sequence parity tool

Dhruv Garg
and
claude
committed
plots: semantic participation heatmap, response-lateness CDF, aggregate time-split, acc-gain dual-axis

Dhruv Garg
and
claude
committed
sync aggregator: guard channel.ends() returning None in sim mode

Dhruv Garg
and
claude
committed
Add handoff doc: trainer overhead, real/sim parity, plots (state + next steps)

Dhruv Garg
and
claude
committed

Commits on Jun 5, 2026

real/sim parity across all 6 baselines + n300 overnight sim infra + plots

Dhruv Garg
and
claude
committed
overnight scripts: portable conda activation + derive repo paths from script location

Dhruv Garg
and
claude
committed
overrun debug: WALL_SEND_TS + MQTT lag decomposition + debug_run.sh

Dhruv Garg
and
claude
committed
scripts: move run logs from /tmp to repo run_logs/ dir

Dhruv Garg
and
claude
committed
scripts: revert log dir back to /tmp

Dhruv Garg
and
claude
committed
timing: 6-component LAG_DECOMP + TIMING_OVERRUN_AGG + wall-clock failsafe

Dhruv Garg
and
claude
committed

Commits on Jun 6, 2026

timing+cpupin+vclock: per-phase trainer timing, CPU pinning, sim clock fix

Dhruv Garg
and
claude
committed
analysis: round-2 fixes — sim_rate, sim_wall_ceiling, pinning+dynamics analysis

Dhruv Garg
and
claude
committed
longer 3h debug run
Dhruv Garg
committed

Commits on Jun 7, 2026

parity: new scripts/parity/ package + K2/K3/K4/K8/K10 clock checks

Dhruv Garg
and
claude
committed

Commits on Jun 8, 2026

cleanup: remove stale artifacts, supersede old parity scripts, trim comments

Dhruv Garg
and
claude
committed