#!/usr/bin/env bash # TRACE study main entry. Run ONE of these per node (within ~30 min of each # other; the nodes wait for one another at a barrier): # # ./run_study.sh check ONLY the fail-fast connectivity pass # ./run_study.sh overnight the time-boxed study (schedule_5node_overnight.json) # ./run_study.sh day2 one-start day2 study (~18h35m); # validates itself before continuing # ./run_study.sh day2-test internal startup-validation schedule # ./run_study.sh full study (schedule_3node.json) # ./run_study.sh 5node full study incl. the H200 pair # ./run_study.sh test ~10 min end-to-end test, tiny jobs # ./run_study.sh local single-node 4-GPU contrast set # ./run_study.sh explicit schedule # # Day2 uses an integrated five-node startup validation. Other modes retain the # standalone connectivity pass; set SKIP_NETCHECK=1 to bypass that older pass. # # Needs: ./setup.sh done, and nodes.conf filled in (copy nodes.conf.example). # Produces: trace-study-result--