ywkang/kernbench2
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
sccl-distributed-allreduce
kernbench2/docs/diagrams/pe_dma_perf/summary.csv
T
ywkang a76487ca48 PE_DMA perf: SIP-wide scenarios + dual outputs + clearer naming 
User asked to surface system-wide congestion (more accurate than
single-cube), bring back the latency-breakdown plot under a separate
filename, and rename the obscure ``streaming`` category.

Scenarios:
  Renamed all_pe_to_pe0 → all_pe_cube0_to_pe0 (clarify cube scope).
  Added two SIP-wide scenarios:
    sip_local_all     — every PE in sip0 (128 total) accesses its own
                        local slice. All paths disjoint (each PE owns
                        its own hbm_ctrl.peX), so the model should
                        scale linearly with cube count.
    sip_hotspot_pe0   — every PE in sip0 (128 total) targets
                        sip0.cube0.pe0_slice. Worst-case hotspot:
                        UCIe inbound + r0c0→hbm_ctrl.pe0 saturated.
  Each bar now carries an ``N=...`` annotation showing the issuer
  count, and the chart titles say the scope explicitly.

Effective BW + util at 16 KB:
  sip_local_all       N=128  eff= 27.2 TB/s  util_a= 83 %
  sip_hotspot_pe0     N=128  eff= 134 GB/s   util_a= 93 %
                                              (UCIe-into-cube0 saturated)

Plots:
  no_congestion.png + congestion.png        — Effective BW utilization
                                              (two bars: single vs aggregate peak)
  breakdown_no_congestion.png +
  breakdown_congestion.png                  — stacked latency breakdown
                                              (renamed from previous)
  summary.csv with columns for both views.

The visual y-cap on BW utilization is 150 %. Bars exceeding it (e.g.
sip_local_all's util_single = 10,639 %) are drawn at the cap with an
upward arrow and the real value annotated. The verification rule for
``util_single`` is loosened to ``≤ n_issuers × 100 % + 5 %`` so
massively-parallel disjoint scenarios pass.

Category renamed: ``streaming`` → ``wire_transfer``. It is the
bulk-transfer time = (n_flits − 1) × flit_bytes / bottleneck_bw — the
cost of streaming the rest of the payload through the slowest wire
after the first flit has arrived.

All checks PASS.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-15 09:43:09 -07:00

34 lines
5.0 KiB
CSV
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
graph,scenario,label,nbytes,n_issuers,total_ns,makespan_ns,min_lat_ns,peak_single_bw_gbs,peak_aggregate_bw_gbs,effective_bw_gbs,util_single_pct,util_aggregate_pct,pe_setup,noc_mesh,ucie,fabric,wire_transfer,hbm_ctrl,contention,path,first_path
no_congestion,local,"SAME_CUBE
PE_LOCAL",16384,1,77.0,,,256.0,256.0,212.7792207792208,83.11688311688312,83.11688311688312,1.0,2.0,0.0,0.0,63.0,9.0,2.0,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0,
no_congestion,same_cube_best,"SAME_CUBE
REMOTE_BEST
(pe0→pe1)",16384,1,82.06,,,256.0,256.0,199.6587862539605,77.99171338045332,77.99171338045332,1.0,5.03,0.0,0.0,63.0,9.0,4.030000000000001,pe0.pe_dma -> cube0.r0c0 -> cube0.r0c1 -> hbm_ctrl.pe1,
no_congestion,same_cube_worst,"SAME_CUBE
REMOTE_WORST
(pe0→pe7)",16384,1,117.50000000000001,,,256.0,256.0,139.4382978723404,54.46808510638297,54.46808510638297,1.0,26.25,0.0,0.0,63.0,9.0,18.250000000000014,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)",16384,1,202.51999999999998,,,128.0,128.0,80.90065178747778,63.20363420896702,63.20363420896702,1.0,6.0,32.510000000000005,0.0,126.0,9.0,28.00999999999999,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)",16384,1,573.1199999999999,,,128.0,128.0,28.587381351200452,22.333891680625353,22.333891680625353,1.0,30.0,219.05999999999995,0.0,126.0,9.0,188.05999999999995,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)",16384,1,408.5216666666663,,,128.0,128.0,40.10558395515541,31.332487464965165,31.332487464965165,1.0,4.0,37.040000000000006,22.09666666666667,126.0,9.0,209.38499999999962,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,"cube0
1×PE → pe0_slice",16384,1,,82.06,82.06,256.0,256.0,199.6587862539605,77.99171338045332,77.99171338045332,1.0,5.03,0.0,0.0,63.0,9.0,4.030000000000001,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
congestion,ctrl_hot_2,"cube0
2×PE → pe0_slice",16384,2,,158.3450000000001,134.2400000000001,256.0,256.0,206.94054122327813,80.83614891534302,80.83614891534302,1.0,5.03,0.0,0.0,63.0,9.0,80.31500000000011,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
congestion,ctrl_hot_3,"cube0
3×PE → pe0_slice",16384,3,,230.0750000000001,139.94000000000008,256.0,256.0,213.6346843420623,83.45104857111808,83.45104857111808,1.0,5.03,0.0,0.0,63.0,9.0,152.0450000000001,,pe1.pe_dma -> cube0.r0c1 -> cube0.r0c0 -> hbm_ctrl.pe0
congestion,ucie_eastbound,"cube0
8×PE corresp.
→ cube1",16384,8,,962.52,438.52,128.0,159.99999999999997,136.17587167019906,106.387399742343,85.10991979387443,1.0,6.0,32.510000000000005,0.0,126.0,9.0,788.01,,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_cube0_to_pe0,"cube0
8×PE → pe0_slice",16384,8,,558.2499999999998,195.0,256.0,256.0,234.7908643081058,91.71518137035383,91.71518137035383,1.0,2.0,0.0,0.0,63.0,9.0,483.2499999999998,,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0
congestion,sip_local_all,"sip0
128×PE → own slice",16384,128,,77.0,77.0,256.0,32768.0,27235.74025974026,10638.961038961039,83.11688311688312,1.0,2.0,0.0,0.0,63.0,9.0,2.0,,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0
congestion,sip_hotspot_pe0,"sip0
128×PE → cube0.pe0_slice",16384,128,,15618.595000000001,204.0,256.0,143.9999999999998,134.2727690935068,52.4503004271511,93.24497853715764,1.0,2.0,0.0,0.0,63.0,9.0,15543.595000000001,,pe0.pe_dma -> cube0.r0c0 -> hbm_ctrl.pe0
Reference in New Issue View Git Blame Copy Permalink