""" gui/pages/explorer.py — ⊞ Run Explorer Render function: render(ctx) ctx keys: ov, runs, tax, lin, age, avg_lin_j, avg_age_j, tax_mult, plan_ms, exec_ms, synth_ms, plan_pct, exec_pct, synth_pct """ import subprocess import streamlit as st import pandas as pd import plotly.express as px import plotly.graph_objects as go from plotly.subplots import make_subplots from gui.config import PROJECT_ROOT, DB_PATH, LIVE_API, WF_COLORS, PL from gui.db import q, q_safe, q1 from gui.helpers import fl, _human_energy, _human_water, _human_carbon, _gauge_html, _bar_gauge_html try: import requests as _req _REQUESTS_OK = True except ImportError: _REQUESTS_OK = False class _req: @staticmethod def get(*a, **kw): raise RuntimeError("requests not installed") try: import yaml as _yaml _YAML_OK = True except ImportError: _YAML_OK = False def render(ctx: dict): ov = ctx["ov"] runs = ctx["runs"] tax = ctx["tax"] avg_lin_j = ctx["avg_lin_j"] avg_age_j = ctx["avg_age_j"] tax_mult = ctx["tax_mult"] plan_ms = ctx["plan_ms"] exec_ms = ctx["exec_ms"] synth_ms = ctx["synth_ms"] plan_pct = ctx["plan_pct"] exec_pct = ctx["exec_pct"] synth_pct = ctx["synth_pct"] lin = ctx["lin"] age = ctx["age"] st.title("Sample Explorer") st.caption("100Hz RAPL energy · CPU c-states · interrupt timeseries per run") if runs.empty: st.info("No runs available — run experiments first.") else: # ── Run picker ──────────────────────────────────────────────────────── _sel = runs[runs.energy_j.notna()].copy() def _lbl(r): task = str(r.task_name or "?")[:22] return (f"Run {int(r.run_id):>4} {r.workflow_type:<8} " f"{r.provider:<6} {r.energy_j:.3f}J {task}") _labels = [_lbl(r) for _, r in _sel.iterrows()] _ids = _sel.run_id.tolist() col_pick, col_stats = st.columns([2, 1]) with col_pick: chosen = st.selectbox("Select run", _labels, help="Sorted newest-first. Shows run_id, workflow, provider, energy, task.") rid = int(_ids[_labels.index(chosen)]) # ── Load all sample tables for this run ─────────────────────────────── _err_ph = st.empty() def load_samples(run_id: int): errors = [] # energy_samples — matches exact DDL es_df, e1 = q_safe(f""" SELECT (timestamp_ns - MIN(timestamp_ns) OVER (PARTITION BY run_id)) / 1e6 AS elapsed_ms, pkg_energy_uj / 1e6 AS pkg_j, core_energy_uj / 1e6 AS core_j, COALESCE(uncore_energy_uj, 0) / 1e6 AS uncore_j, COALESCE(dram_energy_uj, 0) / 1e6 AS dram_j FROM energy_samples WHERE run_id = {run_id} ORDER BY timestamp_ns """) if e1: errors.append(f"energy_samples: {e1}") # cpu_samples — matches exact DDL (no 'ipc' column in cpu_samples — it IS there) cs_df, e2 = q_safe(f""" SELECT (timestamp_ns - MIN(timestamp_ns) OVER (PARTITION BY run_id)) / 1e6 AS elapsed_ms, COALESCE(cpu_util_percent, 0) AS cpu_util_percent, COALESCE(ipc, 0) AS ipc, COALESCE(package_power, 0) AS pkg_w, COALESCE(dram_power, 0) AS dram_w, COALESCE(c1_residency, 0) AS c1, COALESCE(c2_residency, 0) AS c2, COALESCE(c3_residency, 0) AS c3, COALESCE(c6_residency, 0) AS c6, COALESCE(c7_residency, 0) AS c7, COALESCE(pkg_c8_residency, 0) AS c8, COALESCE(package_temp, 0) AS pkg_temp FROM cpu_samples WHERE run_id = {run_id} ORDER BY timestamp_ns """) if e2: errors.append(f"cpu_samples: {e2}") # interrupt_samples — matches exact DDL irq_df, e3 = q_safe(f""" SELECT (timestamp_ns - MIN(timestamp_ns) OVER (PARTITION BY run_id)) / 1e6 AS elapsed_ms, interrupts_per_sec FROM interrupt_samples WHERE run_id = {run_id} ORDER BY timestamp_ns """) if e3: errors.append(f"interrupt_samples: {e3}") # orchestration_events (nullable table — may be empty for linear runs) ev_df, e4 = q_safe(f""" SELECT (start_time_ns - MIN(start_time_ns) OVER (PARTITION BY run_id)) / 1e6 AS start_ms, duration_ns / 1e6 AS duration_ms, phase, event_type FROM orchestration_events WHERE run_id = {run_id} ORDER BY start_time_ns """) if e4: errors.append(f"orchestration_events: {e4}") return es_df, cs_df, irq_df, ev_df, errors with st.spinner(f"Loading samples for run {rid}…"): es, cs, irq, ev, errs = load_samples(rid) if errs: for err in errs: st.error(f"⚠ SQL error — {err}") # ── KPI row ─────────────────────────────────────────────────────────── _r = runs[runs.run_id == rid].iloc[0] k1,k2,k3,k4,k5 = st.columns(5) k1.metric("Run", f"{rid} — {_r.workflow_type}") k2.metric("Total energy", f"{_r.energy_j:.3f}J") k3.metric("Energy samples", f"{len(es):,}") k4.metric("CPU samples", f"{len(cs):,}") k5.metric("Interrupt samples",f"{len(irq):,}") st.divider() # ── Power timeseries ────────────────────────────────────────────────── if not es.empty and len(es) > 2: _es = es.copy() dt = (_es.elapsed_ms.diff() / 1000).replace(0, float("nan")) # sec, avoid /0 _es["pkg_w"] = (_es.pkg_j.diff() / dt).clip(lower=0) _es["core_w"] = (_es.core_j.diff() / dt).clip(lower=0) _es["dram_w"] = (_es.dram_j.diff() / dt).clip(lower=0) _es = _es.iloc[1:].copy() # drop first NaN row col1, col2 = st.columns(2) with col1: st.markdown("**⚡ Power over time (Watts)**") st.caption("Instantaneous power derived from RAPL Δenergy / Δtime at 100Hz") fig = go.Figure() fig.add_trace(go.Scatter( x=_es.elapsed_ms, y=_es.pkg_w, name="PKG", line=dict(color="#3b82f6", width=1.5), fill="tozeroy", fillcolor="rgba(59,130,246,.08)")) fig.add_trace(go.Scatter( x=_es.elapsed_ms, y=_es.core_w, name="Core", line=dict(color="#22c55e", width=1))) fig.add_trace(go.Scatter( x=_es.elapsed_ms, y=_es.dram_w, name="DRAM", line=dict(color="#a78bfa", width=1))) st.plotly_chart(fl(fig, xaxis_title="elapsed ms", yaxis_title="Watts"), use_container_width=True) with col2: st.markdown("**∫ Cumulative energy (Joules)**") st.caption("Raw RAPL counter values — monotonically increasing throughout run") fig2 = go.Figure() fig2.add_trace(go.Scatter( x=es.elapsed_ms, y=es.pkg_j, name="PKG", line=dict(color="#3b82f6", width=1.5))) fig2.add_trace(go.Scatter( x=es.elapsed_ms, y=es.core_j, name="Core", line=dict(color="#22c55e", width=1))) fig2.add_trace(go.Scatter( x=es.elapsed_ms, y=es.dram_j, name="DRAM", line=dict(color="#a78bfa", width=1))) st.plotly_chart(fl(fig2, xaxis_title="elapsed ms", yaxis_title="Joules"), use_container_width=True) else: st.info(f"No energy_samples found for run {rid}. " f"Check that the RAPL collector ran during this experiment.") # ── CPU + C-States ──────────────────────────────────────────────────── if not cs.empty: col1, col2 = st.columns(2) with col1: st.markdown("**IPC + CPU utilisation**") st.caption("IPC = instructions/cycle (left axis) · util% (right axis)") fig3 = make_subplots(specs=[[{"secondary_y": True}]]) fig3.add_trace(go.Scatter( x=cs.elapsed_ms, y=cs.ipc, name="IPC", line=dict(color="#22c55e", width=1.5)), secondary_y=False) fig3.add_trace(go.Scatter( x=cs.elapsed_ms, y=cs.cpu_util_percent, name="CPU Util%", line=dict(color="#f59e0b", width=1)), secondary_y=True) fig3.update_layout(**PL) fig3.update_yaxes(title_text="IPC", secondary_y=False, gridcolor="#1e2d45", tickfont=dict(size=9)) fig3.update_yaxes(title_text="CPU Util%", secondary_y=True, gridcolor="rgba(0,0,0,0)", tickfont=dict(size=9)) st.plotly_chart(fig3, use_container_width=True) with col2: st.markdown("**C-State residency over time**") st.caption("Stacked: C6/C7 = deep sleep, C0 = active execution") fig4 = go.Figure() for _col, _color, _name in [ ("c7","#f59e0b","C7 (deepest)"), ("c6","#22c55e","C6"), ("c3","#38bdf8","C3"), ("c2","#3b82f6","C2"), ("c1","#a78bfa","C1"), ]: fig4.add_trace(go.Scatter( x=cs.elapsed_ms, y=cs[_col], name=_name, line=dict(color=_color, width=1), stackgroup="cstate", fill="tonexty")) st.plotly_chart(fl(fig4, xaxis_title="elapsed ms", yaxis_title="Residency %"), use_container_width=True) # Package power + temp side by side col3, col4 = st.columns(2) with col3: st.markdown("**Package power (W) from turbostat**") st.caption("Direct turbostat reading — cross-check vs RAPL power above") fig5 = go.Figure() fig5.add_trace(go.Scatter( x=cs.elapsed_ms, y=cs.pkg_w, name="PKG W", line=dict(color="#3b82f6", width=1.5), fill="tozeroy", fillcolor="rgba(59,130,246,.06)")) if cs.dram_w.any(): fig5.add_trace(go.Scatter( x=cs.elapsed_ms, y=cs.dram_w, name="DRAM W", line=dict(color="#a78bfa", width=1))) st.plotly_chart(fl(fig5, xaxis_title="elapsed ms", yaxis_title="Watts"), use_container_width=True) with col4: if cs.pkg_temp.any(): st.markdown("**Package temperature (°C)**") st.caption("Thermal headroom — sustained load shows gradual rise") fig6 = go.Figure() fig6.add_trace(go.Scatter( x=cs.elapsed_ms, y=cs.pkg_temp, name="Temp °C", line=dict(color="#ef4444", width=1.5), fill="tozeroy", fillcolor="rgba(239,68,68,.06)")) st.plotly_chart(fl(fig6, xaxis_title="elapsed ms", yaxis_title="°C"), use_container_width=True) else: st.info("No temperature data for this run.") # ── IRQ timeseries ──────────────────────────────────────────────────── if not irq.empty: st.markdown("**IRQ rate (interrupts/sec)**") st.caption("Spikes = API response arrivals or timer interrupts during phase transitions") fig7 = go.Figure() fig7.add_trace(go.Scatter( x=irq.elapsed_ms, y=irq.interrupts_per_sec, name="IRQ/s", line=dict(color="#ef4444", width=1.5), fill="tozeroy", fillcolor="rgba(239,68,68,.06)")) st.plotly_chart(fl(fig7, xaxis_title="elapsed ms", yaxis_title="IRQ/s"), use_container_width=True) # ── Orchestration events Gantt ──────────────────────────────────────── if not ev.empty: st.markdown("**Orchestration Event Timeline**") st.caption("Each bar = one agent phase. Width = duration. Hover for detail.") PHASE_C = {"planning":"#f59e0b","execution":"#3b82f6","synthesis":"#a78bfa"} fig8 = go.Figure() for _, row in ev.iterrows(): fig8.add_trace(go.Bar( x=[row.duration_ms], y=[f"{row.phase or '?'} / {row.event_type or '?'}"], base=row.start_ms, orientation="h", marker_color=PHASE_C.get(str(row.phase), "#3b82f6"), hovertemplate=(f"{row.event_type}
" f"Phase: {row.phase}
" f"Duration: {row.duration_ms:.0f}ms"), )) fig8.update_layout(**PL, xaxis_title="elapsed ms", showlegend=False, height=max(200, len(ev) * 32)) st.plotly_chart(fig8, use_container_width=True) elif _r.workflow_type == "agentic": st.info("No orchestration events recorded for this agentic run.") # ══════════════════════════════════════════════════════════════════════════════ # EXPERIMENTS # ══════════════════════════════════════════════════════════════════════════════