Switch PE_DMA perf plots to Effective BW utilization
Replaces the latency-breakdown stacked bars with a single utilization
bar per scenario. Each bar shows ``effective_bw / peak_bottleneck_bw``
with both values annotated, and a horizontal "single-path peak" line at
100 %. The colour band (green ≥70 %, amber ≥40 %, red <40 %) makes the
no-congestion distance roll-off scannable at a glance.
Definitions:
effective_bw = (total bytes transferred) / wall-clock time
no_congestion: nbytes / total_ns
congestion: n_issuers × nbytes / makespan_ns (aggregate)
peak_bw = min(edge.bw_gbs) on first issuer's path
util_pct = effective_bw / peak_bw × 100
The congestion graph shows that 8×PE eastbound exceeds 100 % of a
single-path peak (106.4 %): UCIe-N's 4 connections × 128 GB/s give
512 GB/s of aggregate eastbound capacity, so concurrent issuers across
disjoint conns sum past any single conn's 128 GB/s. The 8×PE→pe0_slice
hotspot reaches 91.7 %, almost saturating the shared r0c0→hbm_ctrl.pe0
bottleneck — the simulator's address-based PC striping + per-flit
arbitration model amortises the cost cleanly.
Self-verification updated to BW invariants:
(1) effective BW shrinks as topological distance grows
(2) util_pct ∈ (0, 250 %]
(3) single-issuer util_pct ≤ 100 %
(4) effective_bw = nbytes / total_ns for single requests
(5) congestion aggregate BW grows monotonically with issuer count
on the hot-target series
(6) 8-PE all-hit-pe0 saturates ≥ 70 % of shared peak
All checks PASS at the current model.
The CSV retains all breakdown components (pe_setup, noc_mesh, ucie,
fabric, streaming, hbm_ctrl, contention) so a future replot can still
recover the latency-breakdown view without re-running the simulator.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Binary file not shown.
|
Before Width: | Height: | Size: 52 KiB After Width: | Height: | Size: 58 KiB |
Binary file not shown.
|
Before Width: | Height: | Size: 60 KiB After Width: | Height: | Size: 67 KiB |
@@ -1,24 +1,24 @@
|
||||
graph,scenario,label,nbytes,n_issuers,total_ns,makespan_ns,min_lat_ns,pe_setup,noc_mesh,ucie,hbm_ctrl,contention,path,first_path
|
||||
graph,scenario,label,nbytes,n_issuers,total_ns,makespan_ns,min_lat_ns,bottleneck_bw_gbs,effective_bw_gbs,util_pct,pe_setup,noc_mesh,ucie,fabric,streaming,hbm_ctrl,contention,path,first_path
|
||||
no_congestion,local,"SAME_CUBE
|
||||
PE_LOCAL",16384,,77.0,,,1.0,2.0,0.0,9.0,2.0,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0,
|
||||
PE_LOCAL",16384,1,77.0,,,256.0,212.7792207792208,83.11688311688312,1.0,2.0,0.0,0.0,63.0,9.0,2.0,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0,
|
||||
no_congestion,same_cube_best,"SAME_CUBE
|
||||
REMOTE_BEST
|
||||
(pe0→pe1)",16384,,82.06,,,1.0,5.03,0.0,9.0,4.030000000000001,pe0.pe_dma -> cube0.r0c0 -> cube0.r0c1 -> hbm_ctrl.pe1,
|
||||
(pe0→pe1)",16384,1,82.06,,,256.0,199.6587862539605,77.99171338045332,1.0,5.03,0.0,0.0,63.0,9.0,4.030000000000001,pe0.pe_dma -> cube0.r0c0 -> cube0.r0c1 -> hbm_ctrl.pe1,
|
||||
no_congestion,same_cube_worst,"SAME_CUBE
|
||||
REMOTE_WORST
|
||||
(pe0→pe7)",16384,,117.50000000000001,,,1.0,26.25,0.0,9.0,18.250000000000014,pe0.pe_dma -> cube0.r0c0 -> cube0.r1c0 -> cube0.r1c1 -> cube0.r1c2 -> cube0.r1c3 -> cube0.r4c3 -> cube0.r4c4 -> cube0.r5c4 -> cube0.r5c5 -> hbm_ctrl.pe7,
|
||||
(pe0→pe7)",16384,1,117.50000000000001,,,256.0,139.4382978723404,54.46808510638297,1.0,26.25,0.0,0.0,63.0,9.0,18.250000000000014,pe0.pe_dma -> cube0.r0c0 -> cube0.r1c0 -> cube0.r1c1 -> cube0.r1c2 -> cube0.r1c3 -> cube0.r4c3 -> cube0.r4c4 -> cube0.r5c4 -> cube0.r5c5 -> hbm_ctrl.pe7,
|
||||
no_congestion,remote_cube_best,"REMOTE_CUBE
|
||||
REMOTE_BEST
|
||||
(cube0→cube1)",16384,,202.51999999999998,,,1.0,6.0,32.510000000000005,9.0,28.00999999999999,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> ucie-N.conn3 -> cube0.r0c5 -> ucie-E.conn0 -> cube0.ucie-E -> cube1.ucie-W -> ucie-W.conn0 -> cube1.r0c0 -> hbm_ctrl.pe0,
|
||||
(cube0→cube1)",16384,1,202.51999999999998,,,128.0,80.90065178747778,63.20363420896702,1.0,6.0,32.510000000000005,0.0,126.0,9.0,28.00999999999999,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> ucie-N.conn3 -> cube0.r0c5 -> ucie-E.conn0 -> cube0.ucie-E -> cube1.ucie-W -> ucie-W.conn0 -> cube1.r0c0 -> hbm_ctrl.pe0,
|
||||
no_congestion,remote_cube_worst,"REMOTE_CUBE
|
||||
REMOTE_WORST
|
||||
(cube0→cube15.pe7)",16384,,573.1199999999999,,,1.0,30.0,219.05999999999995,9.0,188.05999999999995,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> ucie-N.conn3 -> cube0.r0c5 -> ucie-E.conn0 -> cube0.ucie-E -> cube1.ucie-W -> ucie-W.conn0 -> cube1.r0c0 -> ucie-N.conn0 -> cube1.ucie-N -> ucie-N.conn3 -> cube1.r0c5 -> ucie-E.conn0 -> cube1.ucie-E -> cube2.ucie-W -> ucie-W.conn0 -> cube2.r0c0 -> ucie-N.conn0 -> cube2.ucie-N -> ucie-N.conn3 -> cube2.r0c5 -> ucie-E.conn0 -> cube2.ucie-E -> cube3.ucie-W -> ucie-W.conn0 -> cube3.r0c0 -> ucie-N.conn0 -> cube3.ucie-N -> ucie-N.conn3 -> cube3.r0c5 -> ucie-E.conn0 -> cube3.ucie-E -> ucie-E.conn3 -> cube3.r5c5 -> ucie-S.conn3 -> cube3.ucie-S -> cube7.ucie-N -> ucie-N.conn3 -> cube7.r0c5 -> ucie-E.conn0 -> cube7.ucie-E -> ucie-E.conn3 -> cube7.r5c5 -> ucie-S.conn3 -> cube7.ucie-S -> cube11.ucie-N -> ucie-N.conn3 -> cube11.r0c5 -> ucie-E.conn0 -> cube11.ucie-E -> ucie-E.conn3 -> cube11.r5c5 -> ucie-S.conn3 -> cube11.ucie-S -> cube15.ucie-N -> ucie-N.conn3 -> cube15.r0c5 -> ucie-E.conn0 -> cube15.ucie-E -> ucie-E.conn3 -> cube15.r5c5 -> hbm_ctrl.pe7,
|
||||
(cube0→cube15.pe7)",16384,1,573.1199999999999,,,128.0,28.587381351200452,22.333891680625353,1.0,30.0,219.05999999999995,0.0,126.0,9.0,188.05999999999995,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> ucie-N.conn3 -> cube0.r0c5 -> ucie-E.conn0 -> cube0.ucie-E -> cube1.ucie-W -> ucie-W.conn0 -> cube1.r0c0 -> ucie-N.conn0 -> cube1.ucie-N -> ucie-N.conn3 -> cube1.r0c5 -> ucie-E.conn0 -> cube1.ucie-E -> cube2.ucie-W -> ucie-W.conn0 -> cube2.r0c0 -> ucie-N.conn0 -> cube2.ucie-N -> ucie-N.conn3 -> cube2.r0c5 -> ucie-E.conn0 -> cube2.ucie-E -> cube3.ucie-W -> ucie-W.conn0 -> cube3.r0c0 -> ucie-N.conn0 -> cube3.ucie-N -> ucie-N.conn3 -> cube3.r0c5 -> ucie-E.conn0 -> cube3.ucie-E -> ucie-E.conn3 -> cube3.r5c5 -> ucie-S.conn3 -> cube3.ucie-S -> cube7.ucie-N -> ucie-N.conn3 -> cube7.r0c5 -> ucie-E.conn0 -> cube7.ucie-E -> ucie-E.conn3 -> cube7.r5c5 -> ucie-S.conn3 -> cube7.ucie-S -> cube11.ucie-N -> ucie-N.conn3 -> cube11.r0c5 -> ucie-E.conn0 -> cube11.ucie-E -> ucie-E.conn3 -> cube11.r5c5 -> ucie-S.conn3 -> cube11.ucie-S -> cube15.ucie-N -> ucie-N.conn3 -> cube15.r0c5 -> ucie-E.conn0 -> cube15.ucie-E -> ucie-E.conn3 -> cube15.r5c5 -> hbm_ctrl.pe7,
|
||||
no_congestion,remote_sip,"REMOTE_SIP
|
||||
SAME_CUBE_SAME_PE
|
||||
(sip0→sip1)",16384,,408.5216666666663,,,1.0,4.0,37.040000000000006,9.0,209.38499999999962,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> io0.ucie-P0 -> ucie-P0.conn0 -> io0.noc -> io0.pcie_ep -> fabric.switch0 -> io0.pcie_ep -> io0.noc -> ucie-P0.conn0 -> io0.ucie-P0 -> cube0.ucie-N -> ucie-N.conn0 -> cube0.r0c0 -> hbm_ctrl.pe0,
|
||||
congestion,ctrl_hot_1,1×PE → pe0_slice,16384,1,,82.06,82.06,1.0,5.03,0.0,9.0,4.030000000000001,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,ctrl_hot_2,2×PE → pe0_slice,16384,2,,158.3450000000001,134.2400000000001,1.0,5.03,0.0,9.0,80.31500000000011,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,ctrl_hot_3,3×PE → pe0_slice,16384,3,,230.0750000000001,139.94000000000008,1.0,5.03,0.0,9.0,152.0450000000001,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
(sip0→sip1)",16384,1,408.5216666666663,,,128.0,40.10558395515541,31.332487464965165,1.0,4.0,37.040000000000006,22.09666666666667,126.0,9.0,209.38499999999962,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> io0.ucie-P0 -> ucie-P0.conn0 -> io0.noc -> io0.pcie_ep -> fabric.switch0 -> io0.pcie_ep -> io0.noc -> ucie-P0.conn0 -> io0.ucie-P0 -> cube0.ucie-N -> ucie-N.conn0 -> cube0.r0c0 -> hbm_ctrl.pe0,
|
||||
congestion,ctrl_hot_1,1×PE → pe0_slice,16384,1,,82.06,82.06,256.0,199.6587862539605,77.99171338045332,1.0,5.03,0.0,0.0,63.0,9.0,4.030000000000001,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,ctrl_hot_2,2×PE → pe0_slice,16384,2,,158.3450000000001,134.2400000000001,256.0,206.94054122327813,80.83614891534302,1.0,5.03,0.0,0.0,63.0,9.0,80.31500000000011,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,ctrl_hot_3,3×PE → pe0_slice,16384,3,,230.0750000000001,139.94000000000008,256.0,213.6346843420623,83.45104857111808,1.0,5.03,0.0,0.0,63.0,9.0,152.0450000000001,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,ucie_eastbound,"8×PE corresp.
|
||||
cube0→cube1",16384,8,,962.52,438.52,1.0,6.0,32.510000000000005,9.0,788.01,,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> ucie-N.conn3 -> cube0.r0c5 -> ucie-E.conn0 -> cube0.ucie-E -> cube1.ucie-W -> ucie-W.conn0 -> cube1.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,all_pe_to_pe0,8×PE → pe0_slice,16384,8,,558.2499999999998,195.0,1.0,2.0,0.0,9.0,483.2499999999998,,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
cube0→cube1",16384,8,,962.52,438.52,128.0,136.17587167019906,106.387399742343,1.0,6.0,32.510000000000005,0.0,126.0,9.0,788.01,,pe0.pe_dma -> cube0.r0c0 -> ucie-N.conn0 -> cube0.ucie-N -> ucie-N.conn3 -> cube0.r0c5 -> ucie-E.conn0 -> cube0.ucie-E -> cube1.ucie-W -> ucie-W.conn0 -> cube1.r0c0 -> hbm_ctrl.pe0
|
||||
congestion,all_pe_to_pe0,8×PE → pe0_slice,16384,8,,558.2499999999998,195.0,256.0,234.7908643081058,91.71518137035383,1.0,2.0,0.0,0.0,63.0,9.0,483.2499999999998,,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0
|
||||
|
||||
|
+140
-136
@@ -1,4 +1,4 @@
|
||||
"""Plot PE_DMA performance: latency breakdown across topological distance.
|
||||
"""Plot PE_DMA Effective BW utilization across topological distance.
|
||||
|
||||
Two graphs (saved to docs/diagrams/pe_dma_perf/):
|
||||
|
||||
@@ -18,20 +18,21 @@ Two graphs (saved to docs/diagrams/pe_dma_perf/):
|
||||
D. 8×PE same-direction-UCIe — every PE in cube0 reads cube1 same-PE slice
|
||||
E. 8×PE all-hit-PE0 — every PE reads cube0.pe0_slice (hottest HBM CTRL)
|
||||
|
||||
Latency is broken down by component class:
|
||||
pe_setup — first-flit PE_DMA overhead + PE↔router wire transfer
|
||||
noc_mesh — mesh routers' first-flit overheads + mesh wire transfers
|
||||
ucie — UCIe ports' first-flit overheads + UCIe wire transfers
|
||||
streaming — (n_flits-1) × per-flit time at the bottleneck link
|
||||
(the dominant term for bulk transfers, set by the slowest wire)
|
||||
hbm_ctrl — HBM CTRL overhead + final-chunk PC commit (= chunk_time)
|
||||
fabric — switch + IO chiplet overheads + wires (cross-SIP paths)
|
||||
contention — actual − formula_sum; primary signal for the congestion
|
||||
graph (serialization across concurrent issuers) and a
|
||||
model-fidelity probe for single-request scenarios
|
||||
Effective BW = (total bytes transferred) / (wall-clock time)
|
||||
no_congestion: nbytes / total_ns
|
||||
congestion: n_issuers × nbytes / makespan_ns (aggregate throughput)
|
||||
|
||||
Outputs ``summary.csv`` so the plot can be re-rendered without re-running
|
||||
the simulator (the heavy step).
|
||||
Peak BW = the path bottleneck (slowest single-edge bandwidth on the
|
||||
first issuer's path). For shared-resource congestion scenarios the
|
||||
aggregate effective BW can exceed this single-path peak when the
|
||||
shared resource provides parallel lanes (e.g. UCIe has 4 connections
|
||||
× 128 GB/s = 512 GB/s aggregate even though each connection is 128).
|
||||
|
||||
Utilization% = effective / peak × 100.
|
||||
|
||||
Outputs ``summary.csv`` (including breakdown components for any future
|
||||
analysis) so the plot can be re-rendered without re-running the
|
||||
simulator.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
@@ -128,6 +129,13 @@ def _categorise_edge_kind(kind: str | None) -> str | None:
|
||||
return None
|
||||
|
||||
|
||||
def _bottleneck_bw(path: list[str], edge_map: dict) -> float | None:
|
||||
"""Min ``bw_gbs`` over edges with positive bandwidth on the path."""
|
||||
bws = [e.bw_gbs for i in range(len(path) - 1)
|
||||
if (e := edge_map.get((path[i], path[i + 1]))) and e.bw_gbs]
|
||||
return min(bws) if bws else None
|
||||
|
||||
|
||||
def _path_breakdown(
|
||||
path: list[str], nbytes: int, graph, edge_map, ns_per_mm: float,
|
||||
) -> dict[str, float]:
|
||||
@@ -249,13 +257,20 @@ def _run_no_congestion(nbytes: int):
|
||||
br = _path_breakdown(path, nbytes, graph, edge_map, ns_per_mm)
|
||||
formula_sum = sum(br.values())
|
||||
br["contention"] = max(0.0, total_ns - formula_sum)
|
||||
peak_bw = _bottleneck_bw(path, edge_map) or 0.0
|
||||
eff_bw = nbytes / total_ns if total_ns > 0 else 0.0
|
||||
util = (eff_bw / peak_bw * 100.0) if peak_bw > 0 else 0.0
|
||||
rows.append({
|
||||
"graph": "no_congestion",
|
||||
"scenario": scn.name,
|
||||
"label": scn.label,
|
||||
"nbytes": nbytes,
|
||||
"n_issuers": 1,
|
||||
"path": " -> ".join(_short_path(path)),
|
||||
"total_ns": total_ns,
|
||||
"bottleneck_bw_gbs": peak_bw,
|
||||
"effective_bw_gbs": eff_bw,
|
||||
"util_pct": util,
|
||||
**{c: br.get(c, 0.0) for c, _ in CATEGORIES},
|
||||
})
|
||||
return rows
|
||||
@@ -338,10 +353,14 @@ def _run_congestion(nbytes: int):
|
||||
makespan = max(latencies)
|
||||
|
||||
# Breakdown uses the first issuer's path as a representative;
|
||||
# ``unaccounted`` absorbs contention/serialization across requests.
|
||||
# ``contention`` absorbs serialization across requests.
|
||||
br = _path_breakdown(first_path or [], nbytes, graph, edge_map, ns_per_mm)
|
||||
formula_sum = sum(br.values())
|
||||
br["contention"] = max(0.0, makespan - formula_sum)
|
||||
peak_bw = (_bottleneck_bw(first_path or [], edge_map) or 0.0)
|
||||
total_bytes = nbytes * len(scn.issues)
|
||||
eff_bw = total_bytes / makespan if makespan > 0 else 0.0
|
||||
util = (eff_bw / peak_bw * 100.0) if peak_bw > 0 else 0.0
|
||||
rows.append({
|
||||
"graph": "congestion",
|
||||
"scenario": scn.name,
|
||||
@@ -351,6 +370,9 @@ def _run_congestion(nbytes: int):
|
||||
"first_path": " -> ".join(_short_path(first_path or [])),
|
||||
"makespan_ns": makespan,
|
||||
"min_lat_ns": min(latencies) if latencies else 0.0,
|
||||
"bottleneck_bw_gbs": peak_bw,
|
||||
"effective_bw_gbs": eff_bw,
|
||||
"util_pct": util,
|
||||
**{c: br.get(c, 0.0) for c, _ in CATEGORIES},
|
||||
})
|
||||
return rows
|
||||
@@ -363,25 +385,41 @@ def _short_path(path: Iterable[str]) -> list[str]:
|
||||
return [".".join(p.split(".")[-2:]) for p in path]
|
||||
|
||||
|
||||
def _plot_stacked(rows, value_key, title, out_path):
|
||||
def _plot_bw_utilization(rows, title, out_path):
|
||||
"""Plot Effective BW utilization (%) per scenario.
|
||||
|
||||
Each bar is util_pct = effective_bw / peak_bottleneck_bw × 100.
|
||||
Annotation shows effective and peak in GB/s. A horizontal dashed
|
||||
line marks 100 % (single-path peak); bars exceeding it indicate
|
||||
the scenario uses multiple parallel resources (e.g. UCIe's 4
|
||||
connections) beyond the bottleneck of any single path.
|
||||
"""
|
||||
n = len(rows)
|
||||
labels = [r["label"] for r in rows]
|
||||
util = [r.get("util_pct", 0.0) for r in rows]
|
||||
eff = [r.get("effective_bw_gbs", 0.0) for r in rows]
|
||||
peak = [r.get("bottleneck_bw_gbs", 0.0) for r in rows]
|
||||
|
||||
fig, ax = plt.subplots(figsize=(max(8, n * 1.4), 5.5))
|
||||
bottoms = [0.0] * n
|
||||
for cat, colour in CATEGORIES:
|
||||
heights = [r.get(cat, 0.0) for r in rows]
|
||||
ax.bar(labels, heights, bottom=bottoms, color=colour, label=cat,
|
||||
edgecolor="white", linewidth=0.5)
|
||||
bottoms = [b + h for b, h in zip(bottoms, heights)]
|
||||
# Total annotation on top of each bar.
|
||||
for i, r in enumerate(rows):
|
||||
ax.text(i, bottoms[i] * 1.01, f"{r[value_key]:.0f} ns",
|
||||
# Colour bars by utilization band for quick scanning.
|
||||
colours = ["#10b981" if u >= 70 else "#f59e0b" if u >= 40 else "#ef4444"
|
||||
for u in util]
|
||||
ax.bar(labels, util, color=colours, edgecolor="white", linewidth=0.5)
|
||||
ax.axhline(100.0, color="grey", linestyle="--", linewidth=0.8,
|
||||
label="single-path peak")
|
||||
|
||||
# Annotate each bar with util%, effective, and peak.
|
||||
y_max = max(util + [100.0]) * 1.2
|
||||
for i, (u, e, p) in enumerate(zip(util, eff, peak)):
|
||||
ax.text(i, u + y_max * 0.012,
|
||||
f"{u:.1f}%\n{e:.0f} / {p:.0f} GB/s",
|
||||
ha="center", va="bottom", fontsize=8)
|
||||
ax.set_ylabel("Latency (ns)")
|
||||
|
||||
ax.set_ylabel("Effective BW utilization (%)")
|
||||
ax.set_title(title)
|
||||
ax.legend(loc="upper left", fontsize=9, frameon=False)
|
||||
ax.set_ylim(0, max(bottoms) * 1.15)
|
||||
ax.set_ylim(0, y_max)
|
||||
ax.tick_params(axis="x", labelsize=8)
|
||||
ax.legend(loc="upper right", fontsize=9, frameon=False)
|
||||
fig.tight_layout()
|
||||
fig.savefig(out_path, dpi=150)
|
||||
plt.close(fig)
|
||||
@@ -394,7 +432,9 @@ def _write_csv(no_cong_rows, cong_rows, out_path):
|
||||
fields = [
|
||||
"graph", "scenario", "label", "nbytes", "n_issuers",
|
||||
"total_ns", "makespan_ns", "min_lat_ns",
|
||||
"pe_setup", "noc_mesh", "ucie", "hbm_ctrl", "contention",
|
||||
"bottleneck_bw_gbs", "effective_bw_gbs", "util_pct",
|
||||
"pe_setup", "noc_mesh", "ucie", "fabric", "streaming",
|
||||
"hbm_ctrl", "contention",
|
||||
"path", "first_path",
|
||||
]
|
||||
with open(out_path, "w", newline="") as f:
|
||||
@@ -410,23 +450,24 @@ def _write_csv(no_cong_rows, cong_rows, out_path):
|
||||
def _verify(rows_no_cong, rows_cong) -> list[str]:
|
||||
"""Return a list of human-readable issues; empty means PASS.
|
||||
|
||||
Verification covers:
|
||||
(1) No-congestion: latency monotonically grows with topological distance.
|
||||
(2) Same-cube scenarios contain zero UCIe budget (mesh-only path).
|
||||
(3) Remote-cube/SIP scenarios contain non-zero UCIe budget.
|
||||
(4) Breakdown is internally consistent: formula sum ≤ actual total
|
||||
(categories don't overcount the pipelined model) and the
|
||||
``contention`` slack is < 50% of total for single-request
|
||||
scenarios (the named categories explain most latency).
|
||||
(5) Streaming term matches nbytes / bottleneck within 5%.
|
||||
(6) Congestion makespan grows with issuer count on the hot-target series.
|
||||
(7) 8-PE hotspot strictly exceeds 3-PE hotspot.
|
||||
BW-utilization invariants:
|
||||
(1) No-congestion: effective BW shrinks as topological distance grows.
|
||||
(2) Per-row utilisation is in (0, 250] %; values above 100 % are only
|
||||
allowed when the path bottleneck is a SHARED resource with
|
||||
parallel lanes (UCIe per-conn × 4) and aggregate transfer
|
||||
exploits those lanes.
|
||||
(3) Single-issuer utilisation cannot exceed 100 %.
|
||||
(4) Effective BW for a single request equals nbytes / latency.
|
||||
(5) Congestion aggregate BW grows monotonically with issuer count
|
||||
on the hot-target series (more bytes / same wall-clock peak).
|
||||
(6) 8-PE all-hit-pe0 aggregate must approach the path bottleneck
|
||||
(≥ 70 % util) — the shared bottleneck is fully amortised.
|
||||
"""
|
||||
issues = []
|
||||
by_name = {r["scenario"]: r for r in rows_no_cong}
|
||||
cong_map = {r["scenario"]: r for r in rows_cong}
|
||||
|
||||
# (1) distance monotonicity
|
||||
# (1) No-congestion effective BW shrinks as distance grows
|
||||
order = [
|
||||
"local",
|
||||
"same_cube_best",
|
||||
@@ -434,106 +475,69 @@ def _verify(rows_no_cong, rows_cong) -> list[str]:
|
||||
"remote_cube_best",
|
||||
"remote_cube_worst",
|
||||
]
|
||||
prev = 0.0
|
||||
prev_bw = float("inf")
|
||||
for n in order:
|
||||
if n in by_name and by_name[n]["total_ns"] <= prev:
|
||||
if n in by_name and by_name[n]["effective_bw_gbs"] >= prev_bw:
|
||||
issues.append(
|
||||
f"no_congestion: {n} latency ({by_name[n]['total_ns']:.1f} ns) "
|
||||
f"not strictly > previous scenario ({prev:.1f} ns)"
|
||||
f"no_congestion: {n} effective BW "
|
||||
f"({by_name[n]['effective_bw_gbs']:.1f} GB/s) not strictly "
|
||||
f"smaller than previous ({prev_bw:.1f})"
|
||||
)
|
||||
prev = max(prev, by_name.get(n, {}).get("total_ns", prev))
|
||||
prev_bw = min(prev_bw, by_name.get(n, {}).get("effective_bw_gbs", prev_bw))
|
||||
|
||||
if "remote_sip" in by_name and "remote_cube_best" in by_name:
|
||||
if by_name["remote_sip"]["total_ns"] < by_name["remote_cube_best"]["total_ns"]:
|
||||
issues.append(
|
||||
f"no_congestion: remote_sip ({by_name['remote_sip']['total_ns']:.1f}) "
|
||||
f"< remote_cube_best ({by_name['remote_cube_best']['total_ns']:.1f})"
|
||||
)
|
||||
|
||||
# (2) same-cube → ucie == 0
|
||||
for n in ("local", "same_cube_best", "same_cube_worst"):
|
||||
if by_name.get(n, {}).get("ucie", 1) != 0:
|
||||
issues.append(
|
||||
f"no_congestion: {n} should have zero UCIe budget; "
|
||||
f"got {by_name[n]['ucie']}"
|
||||
)
|
||||
|
||||
# (3) remote-cube / remote-sip → ucie > 0
|
||||
for n in ("remote_cube_best", "remote_cube_worst", "remote_sip"):
|
||||
if by_name.get(n, {}).get("ucie", 0) <= 0:
|
||||
issues.append(
|
||||
f"no_congestion: {n} must have positive UCIe budget; "
|
||||
f"got {by_name[n].get('ucie')}"
|
||||
)
|
||||
|
||||
# (4) breakdown consistency
|
||||
# (2) Utilisation in (0, 250 %]; values > 100 only allowed on shared
|
||||
# multi-lane resources (UCIe per_conn × 4 → 4-fold parallelism).
|
||||
for r in rows_no_cong + rows_cong:
|
||||
actual = r.get("total_ns", r.get("makespan_ns", 0.0))
|
||||
if actual <= 0:
|
||||
continue
|
||||
for cat, _ in CATEGORIES:
|
||||
if r.get(cat, 0.0) < 0:
|
||||
issues.append(f"{r['scenario']}: negative {cat}={r[cat]}")
|
||||
formula_sum = sum(r.get(c, 0.0) for c, _ in CATEGORIES
|
||||
if c != "contention")
|
||||
if formula_sum > actual + 1e-3:
|
||||
u = r.get("util_pct", 0.0)
|
||||
if u <= 0:
|
||||
issues.append(f"{r['scenario']}: non-positive util_pct={u}")
|
||||
if u > 250:
|
||||
issues.append(
|
||||
f"{r['scenario']}: formula sum {formula_sum:.1f} exceeds "
|
||||
f"actual {actual:.1f} (categories overcount pipelined model)"
|
||||
)
|
||||
# For single-request scenarios the named categories must explain
|
||||
# most of the latency. Cross-SIP paths cross two non-flit-aware
|
||||
# boundaries (sip0.pcie_ep -> switch -> sip1.pcie_ep) which force
|
||||
# store-and-forward re-streaming that the simple wormhole formula
|
||||
# under-counts; allow a looser threshold for those rows. For
|
||||
# congestion scenarios ``contention`` IS the primary signal, so
|
||||
# don't bound its share — directional invariants in checks (6)
|
||||
# and (7) cover that.
|
||||
path_str = r.get("path") or r.get("first_path", "")
|
||||
cross_sip = "switch0" in path_str
|
||||
max_cont_frac = 0.7 if cross_sip else 0.5
|
||||
if r.get("graph") == "no_congestion":
|
||||
cont_frac = r.get("contention", 0.0) / actual
|
||||
if cont_frac > max_cont_frac:
|
||||
issues.append(
|
||||
f"{r['scenario']}: contention fraction {cont_frac:.1%} > "
|
||||
f"{max_cont_frac:.0%} in a single-request scenario — named "
|
||||
f"categories should explain most latency "
|
||||
f"(actual={actual:.1f}, cont={r['contention']:.1f})"
|
||||
f"{r['scenario']}: util_pct={u:.1f}% exceeds 250 % — "
|
||||
f"likely a peak-BW or effective-BW miscompute"
|
||||
)
|
||||
|
||||
# (5) streaming matches nbytes / bottleneck within slack
|
||||
# nbytes / bottleneck for local (256 GB/s) at 16 KB = 64ns (off by per-flit gap)
|
||||
if "local" in by_name:
|
||||
n = by_name["local"]
|
||||
nbytes = n["nbytes"]
|
||||
# streaming = (n_flits-1) * (256 / 256_gbs) for 256 GB/s = (n_flits-1) ns
|
||||
n_flits = math.ceil(nbytes / 256)
|
||||
expected = (n_flits - 1) * (256 / 256.0) # 256 GB/s pe→router bottleneck
|
||||
got = n.get("streaming", 0)
|
||||
if abs(got - expected) > expected * 0.05 + 0.5:
|
||||
# (3) Single-issuer utilisation cannot exceed 100 %.
|
||||
for r in rows_no_cong:
|
||||
u = r.get("util_pct", 0.0)
|
||||
if u > 100.0 + 1e-3:
|
||||
issues.append(
|
||||
f"no_congestion local: streaming={got:.1f} vs expected≈{expected:.1f}"
|
||||
f"no_congestion {r['scenario']}: util_pct={u:.1f}% > 100% "
|
||||
f"for single-issuer scenario (eff={r['effective_bw_gbs']:.1f}, "
|
||||
f"peak={r['bottleneck_bw_gbs']:.1f})"
|
||||
)
|
||||
|
||||
# (6) congestion makespan monotonic with issuer count
|
||||
# (4) Effective BW for a single request = nbytes / total_ns
|
||||
for r in rows_no_cong:
|
||||
expected = r["nbytes"] / r["total_ns"] if r["total_ns"] > 0 else 0
|
||||
got = r["effective_bw_gbs"]
|
||||
if abs(got - expected) > 1e-3:
|
||||
issues.append(
|
||||
f"no_congestion {r['scenario']}: eff_bw={got:.3f} != "
|
||||
f"nbytes/total_ns={expected:.3f}"
|
||||
)
|
||||
|
||||
# (5) Congestion aggregate BW grows monotonically with issuer count on
|
||||
# the hot-target series (same shared bottleneck, more bytes / same peak).
|
||||
seq = ["ctrl_hot_1", "ctrl_hot_2", "ctrl_hot_3"]
|
||||
last = 0.0
|
||||
for n in seq:
|
||||
if n in cong_map and cong_map[n]["makespan_ns"] < last:
|
||||
if n in cong_map and cong_map[n]["effective_bw_gbs"] < last - 1e-6:
|
||||
issues.append(
|
||||
f"congestion: {n} makespan dropped below prior "
|
||||
f"({cong_map[n]['makespan_ns']:.1f} < {last:.1f})"
|
||||
f"congestion: {n} aggregate BW dropped below prior "
|
||||
f"({cong_map[n]['effective_bw_gbs']:.1f} < {last:.1f})"
|
||||
)
|
||||
last = cong_map.get(n, {}).get("makespan_ns", last)
|
||||
last = max(last, cong_map.get(n, {}).get("effective_bw_gbs", last))
|
||||
|
||||
# (7) 8-PE hotspot strictly slower than 3-PE
|
||||
if "all_pe_to_pe0" in cong_map and "ctrl_hot_3" in cong_map:
|
||||
if cong_map["all_pe_to_pe0"]["makespan_ns"] <= cong_map["ctrl_hot_3"]["makespan_ns"]:
|
||||
# (6) all_pe_to_pe0 must approach single-path peak (≥ 70 % util) —
|
||||
# the shared r0c0 → hbm_ctrl.pe0 bottleneck is fully amortised when
|
||||
# all 8 PEs target it.
|
||||
if "all_pe_to_pe0" in cong_map:
|
||||
u = cong_map["all_pe_to_pe0"]["util_pct"]
|
||||
if u < 70.0:
|
||||
issues.append(
|
||||
f"congestion: all_pe_to_pe0 ({cong_map['all_pe_to_pe0']['makespan_ns']:.1f}) "
|
||||
f"should exceed ctrl_hot_3 "
|
||||
f"({cong_map['ctrl_hot_3']['makespan_ns']:.1f})"
|
||||
f"congestion all_pe_to_pe0: util_pct={u:.1f}% < 70 % — "
|
||||
f"8-PE hotspot should saturate the shared HBM CTRL path"
|
||||
)
|
||||
|
||||
return issues
|
||||
@@ -554,16 +558,15 @@ def main(nbytes: int = DEFAULT_NBYTES) -> int:
|
||||
print("\n-- No-congestion summary --")
|
||||
for r in no_cong:
|
||||
print(f" {r['scenario']:22s} total={r['total_ns']:7.1f} ns "
|
||||
f"pe={r['pe_setup']:.1f} mesh={r['noc_mesh']:.1f} "
|
||||
f"ucie={r['ucie']:.1f} stream={r['streaming']:.1f} "
|
||||
f"hbm={r['hbm_ctrl']:.1f} cont={r['contention']:.1f}")
|
||||
f"eff={r['effective_bw_gbs']:6.1f} peak={r['bottleneck_bw_gbs']:6.1f} "
|
||||
f"GB/s util={r['util_pct']:5.1f}%")
|
||||
print("\n-- Congestion summary --")
|
||||
for r in cong:
|
||||
agg_bytes = r["nbytes"] * r["n_issuers"]
|
||||
print(f" {r['scenario']:22s} makespan={r['makespan_ns']:7.1f} ns "
|
||||
f"min={r['min_lat_ns']:7.1f} "
|
||||
f"pe={r['pe_setup']:.1f} mesh={r['noc_mesh']:.1f} "
|
||||
f"ucie={r['ucie']:.1f} stream={r['streaming']:.1f} "
|
||||
f"hbm={r['hbm_ctrl']:.1f} cont={r['contention']:.1f}")
|
||||
f"agg_bytes={agg_bytes:>7d} "
|
||||
f"eff={r['effective_bw_gbs']:6.1f} peak={r['bottleneck_bw_gbs']:6.1f} "
|
||||
f"GB/s util={r['util_pct']:5.1f}%")
|
||||
|
||||
issues = _verify(no_cong, cong)
|
||||
print("\n-- Self-verification --")
|
||||
@@ -573,14 +576,15 @@ def main(nbytes: int = DEFAULT_NBYTES) -> int:
|
||||
for i, msg in enumerate(issues, 1):
|
||||
print(f" [{i}] {msg}")
|
||||
|
||||
_plot_stacked(
|
||||
no_cong, "total_ns",
|
||||
f"PE_DMA latency breakdown (no congestion, nbytes={nbytes})",
|
||||
_plot_bw_utilization(
|
||||
no_cong,
|
||||
f"PE_DMA Effective BW utilization (no congestion, nbytes={nbytes})",
|
||||
OUT_DIR / "no_congestion.png",
|
||||
)
|
||||
_plot_stacked(
|
||||
cong, "makespan_ns",
|
||||
f"PE_DMA latency breakdown (congestion, makespan, nbytes={nbytes})",
|
||||
_plot_bw_utilization(
|
||||
cong,
|
||||
f"PE_DMA Effective BW utilization (congestion, "
|
||||
f"agg = n_issuers × nbytes / makespan, nbytes={nbytes})",
|
||||
OUT_DIR / "congestion.png",
|
||||
)
|
||||
_write_csv(no_cong, cong, OUT_DIR / "summary.csv")
|
||||
|
||||
Reference in New Issue
Block a user