Add single-mode dispatch to spec_compression_stress.sh

[YWHyuk]

Summary

Lets serving/spec_compression_stress.sh run a single mode per invocation, so each mode can be fanned out into its own filesystem-isolated container in parallel.

Why

The script ran baseline → self_verify → cpu_verify sequentially in one process. The simulator regenerates astra-sim/inputs/{network.yml, system.json, memory_expansion.json} from the cluster config on every run, so simply backgrounding the three calls would race on those shared files and corrupt the in-flight sim's view of its own network / system / memory config. The cleanest fix is to keep the script unchanged inside, but let the caller run each mode in its own sim docker container — the per-container overlay layer isolates the writes, no code patch needed.

Changes

Invocation	Behaviour
./script.sh (no arg)	Sequential, all three modes — legacy behaviour
./script.sh baseline / self_verify / cpu_verify	Runs only the named mode
./script.sh prepare	Workload synthesis + scaled-config generation only, no sim runs
./script.sh <anything else>	Errors out with usage hint

Parallel dispatch pattern (documented in header)

IMAGE="ghcr.io/psal-postech/llmservingsimspec/sim:latest"
export RUN_ID="$(date +%Y%m%d-%H%M%S)"
OUTPUT_DIR="outputs/spec_compression_stress_$RUN_ID"

# 1. Prepare workload + scaled configs once on the host so the parallel
#    runs share the same inputs and don't race on workload synthesis.
./serving/spec_compression_stress.sh prepare

# 2. Launch each mode in its own container.
for mode in baseline self_verify cpu_verify; do
    docker run --rm --name "stress_${mode}_$$" \
        -v "$(pwd)/configs":/workspace/configs:ro \
        -v "$(pwd)/$OUTPUT_DIR":/workspace/$OUTPUT_DIR \
        -e RUN_ID="$RUN_ID" \
        "$IMAGE" \
        bash /workspace/serving/spec_compression_stress.sh "$mode" \
        > "$OUTPUT_DIR/${mode}.stdout" 2>&1 &
done
wait

Each container gets its own overlay layer so the simulator's astra-sim/inputs/* writes are local to that mode. No code change in the simulator was needed.

Test plan

• ./script.sh (no arg) still runs all three sequentially with identical outputs to before.
• ./script.sh baseline produces just baseline.csv and baseline.log under the output dir.
• ./script.sh prepare produces workload.jsonl (and the scaled configs if H100_SCALE != 1) and exits.
• ./script.sh nonsense errors out with the usage hint and a non-zero exit.
• Parallel-dispatch pattern from the docstring runs three sim containers in parallel against a shared OUTPUT_DIR, producing all three CSVs without astra-sim/inputs/* corruption.

---

Generated by Claude Code