Running times in detailed routing as a function of design parameters
Tool: Detailed Routing
Subcategory: Performance analysis
Conversation
oharboe
mock-array-big has 3 essential parameters that can be adjusted: array area(number of elements), element size (area) and data width (bits).
I thought it would be interesting to see how detailed routing (TritonRoute) varies across these parameters, so I made a python snippet(https://github.com/Pinata-Consulting/OpenROAD-flow-scripts/commits/plot-running-times) to gather some data and plot them
The graph below is interactive(measured on a machine with 16 threads and 64gByte of RAM, where memory consumption appears to peak at 20gByte), which helps in terms of understanding the 4 dimensions being represented:
Video: https://drive.google.com/file/d/1pmkIZh6tUOrWhAkVRgm0fxCD7yaf766Q/view
Snapshot:
8 x 8 64 bit array detailed route times plotted against threads
Conclusion so far
Detailed routing is surprisingly slow when the area is large, but everything else is the same.
This run takes 40 minutes or so, APTOS_MOCK_PITCH_SCALE=20 make FLOW_VARIANT=scale-20 route DESIGN_CONFIG=designs/asap7/mock-array-big/config.mk whereas APTOS_MOCK_PITCH_SCALE=2 make FLOW_VARIANT=scale-2 DESIGN_CONFIG=designs/asap7/mock-array-big/config.mk completes in a few minutes
Detailed of testing...
It looks like the performance of detailed routing is sensitive to area:
MOCK_ARRAY_PITCH_SCALE=2 make OPENROAD_EXE="perf record -o perf.data ~/OpenROAD-flow-scripts/tools/install/OpenROAD/bin/openroad" DESIGN_CONFIG=designs/asap7/mock-array-big/config.mk FLOW_VARIANT=scale-2 route
Samples: 15M of event 'cycles', Event count (approx.): 14285585618565
Overhead Command Shared Object Symbol
16,94% openroad openroad [.] fr::FlexGridGraph::initEdges
8,38% openroad openroad [.] fr::FlexGridGraph::outOfDieVia
7,46% openroad openroad [.] fr::FlexDRWorker::modMinSpacingCostVia
5,92% openroad openroad [.] fr::FlexGCWorker::Impl::isCornerOverlap
4,62% openroad libstdc++.so.6.0.30 [.] std::_Rb_tree_increment
4,23% openroad [unknown] [k] 0xffffffffa7b87f27
4,20% openroad openroad [.] fr::FlexDRWorker::modMinSpacingCostPlanar
3,31% openroad openroad [.] fr::frLayer::getMinSpacingValue
3,22% openroad openroad [.] fr::FlexGridGraph::getIdx
2,93% openroad openroad [.] fr::FlexGCWorker::Impl::initNet_pins_polygonCorners_helper
2,48% openroad openroad [.] fr::FlexDRWorker::modMinSpacingCostVia_eol
2,46% openroad openroad [.] box2boxDistSquareNew
2,37% openroad openroad [.] fr::FlexDRWorker::modBlockedPlanar
2,17% openroad openroad [.] fr::FlexGridGraph::addFixedShapeCostVia
1,71% openroad openroad [.] odb::dbTransform::apply
1,45% openroad openroad [.] fr::FlexGridGraph::isEdgeInBox
1,43% openroad openroad [.] odb::dbTransform::apply
1,28% openroad openroad [.] fr::FlexDRWorker::modBlockedVia
1,25% openroad openroad [.] std::vector<fr::FlexGridGraph::Node, std::allocator<fr::Fle
Same run, but bigger area
MOCK_ARRAY_PITCH_SCALE=20 make OPENROAD_EXE="perf record -o perf.data ~/OpenROAD-flow-scripts/tools/install/OpenROAD/bin/openroad" DESIGN_CONFIG=designs/asap7/mock-array-big/config.mk FLOW_VARIANT=scale-20 route
2_4_floorplan_macro.odb:
Overhead Command Shared Object Symbol
15,25% openroad openroad [.] fr::FlexGridGraph::initEdges
14,39% openroad openroad [.] fr::FlexGCWorker::Impl::isCornerOverlap
7,94% openroad openroad [.] fr::FlexGCWorker::Impl::initNet_pins_polygonCorners_helper
7,67% openroad openroad [.] fr::FlexGridGraph::outOfDieVia
6,08% openroad openroad [.] fr::FlexDRWorker::modMinSpacingCostVia
3,98% openroad libstdc++.so.6.0.30 [.] std::_Rb_tree_increment
3,80% openroad openroad [.] fr::FlexDRWorker::modMinSpacingCostPlanar
2,84% openroad [unknown] [k] 0xffffffffa7b87f27
2,64% openroad openroad [.] fr::FlexGridGraph::getIdx
2,62% openroad openroad [.] fr::frLayer::getMinSpacingValue
2,11% openroad openroad [.] box2boxDistSquareNew
1,66% openroad openroad [.] fr::FlexDRWorker::modMinSpacingCostVia_eol
1,45% openroad openroad [.] odb::dbTransform::apply
1,40% openroad openroad [.] fr::FlexGridGraph::isEdgeInBox
1,36% openroad openroad [.] std::vector<fr::FlexGridGraph::Node, std::allocator<fr::Fle
1,31% openroad openroad [.] fr::FlexGridGraph::addFixedShapeCostVia
1,25% openroad openroad [.] boost::geometry::index::detail::rtree::visitors::spatial_qu
1,20% openroad openroad [.] odb::dbTransform::apply
1,15% openroad openroad [.] fr::FlexDRWorker::modBlockedPlanar
0,88% openroad openroad [.] fr::frSpacingTablePrlConstraint::typeId
0,86% openroad libc.so.6 [.] malloc
It's interesting thing that the 1st run with 1200 violations takes 20 minutes, the 2nd run with 700 violations takes 12 seconds:
[INFO DRT-0195] Start 1st optimization iteration.
Completing 10% with 5729 violations.
elapsed time = 00:02:00, memory = 29854.12 (MB).
Completing 20% with 4647 violations.
elapsed time = 00:04:04, memory = 30005.05 (MB).
Completing 30% with 3871 violations.
elapsed time = 00:06:04, memory = 30034.33 (MB).
Completing 40% with 3694 violations.
elapsed time = 00:08:04, memory = 30100.05 (MB).
Completing 50% with 1899 violations.
elapsed time = 00:10:00, memory = 28813.91 (MB).
Completing 60% with 1471 violations.
elapsed time = 00:11:58, memory = 29945.56 (MB).
Completing 70% with 1042 violations.
elapsed time = 00:13:55, memory = 29941.39 (MB).
Completing 80% with 908 violations.
elapsed time = 00:15:49, memory = 29953.07 (MB).
Completing 90% with 805 violations.
elapsed time = 00:17:44, memory = 30011.33 (MB).
Completing 100% with 738 violations.
elapsed time = 00:19:38, memory = 28749.75 (MB).
[INFO DRT-0199] Number of violations = 1176.
Viol/Layer M2 M3 V3 M4 V4 M5 M6 M7
Corner Spacing 0 0 0 2 0 0 0 0
CutSpcTbl 0 0 5 0 2 0 0 0
EOL 0 0 0 4 0 1 0 0
Metal Spacing 209 105 0 92 0 2 0 0
Recheck 12 0 0 91 0 247 2 86
Short 157 153 0 0 0 0 0 0
eolKeepOut 0 0 0 3 0 3 0 0
[INFO DRT-0267] cpu time = 11:11:56, elapsed time = 00:19:39, memory = 28720.20 (MB), peak = 30179.07 (MB)
Total wire length = 330230 um.
Total wire length on LAYER M1 = 0 um.
Total wire length on LAYER M2 = 32835 um.
Total wire length on LAYER M3 = 28341 um.
Total wire length on LAYER M4 = 130089 um.
Total wire length on LAYER M5 = 91062 um.
Total wire length on LAYER M6 = 30826 um.
Total wire length on LAYER M7 = 17074 um.
Total wire length on LAYER M8 = 0 um.
Total wire length on LAYER M9 = 0 um.
Total wire length on LAYER Pad = 0 um.
Total number of vias = 41385.
Up-via summary (total 41385):.
----------------
Active 0
M1 1424
M2 5855
M3 20852
M4 11321
M5 1354
M6 579
M7 0
M8 0
M9 0
----------------
41385
[INFO DRT-0195] Start 2nd optimization iteration.
Completing 10% with 1154 violations.
elapsed time = 00:00:02, memory = 28720.20 (MB).
Completing 20% with 1135 violations.
elapsed time = 00:00:03, memory = 28720.20 (MB).
Completing 30% with 1116 violations.
elapsed time = 00:00:06, memory = 28720.20 (MB).
Completing 40% with 1078 violations.
elapsed time = 00:00:10, memory = 28716.63 (MB).
Completing 50% with 1059 violations.
elapsed time = 00:00:13, memory = 28716.63 (MB).
Completing 60% with 978 violations.
elapsed time = 00:00:16, memory = 28716.63 (MB).
Completing 70% with 945 violations.
elapsed time = 00:00:17, memory = 28716.63 (MB).
Completing 80% with 920 violations.
elapsed time = 00:00:20, memory = 28716.63 (MB).
Completing 90% with 755 violations.
elapsed time = 00:00:24, memory = 28716.63 (MB).
Completing 100% with 717 violations.
elapsed time = 00:00:27, memory = 28716.63 (MB).
[INFO DRT-0199] Number of violations = 755.
maliberty
How many threads in your experiment?
oharboe
@rovinski I would not have guessed that the area of the design is what really drives up the time of detailed routing.
maliberty
The area of the design will definitely matter. You will have tracks at the same pitch and have to traverse more of them during routing a larger area.
maliberty
Rather than plot "detailed route times plotted against threads" it is usually easier to look at speedup (t(1)/t(N)) vs threads. It makes it clearer how close to ideal scaling you are getting (the 45 degree line is perfect).
oharboe
@maliberty @rovinski I fixed the image for the 1000x1000um when I took a snapshot. I accidentally used the small macro, the spacing between the macros should be very short, nearly abutment.
oharboe
It looks like the expected behavior is that the 0th and the 1st iteration will take a long time for top level designs with lots of wires and little contention, even if these routes are horizontal and vertical and that the low hanging fruit has already been picked.
binsa7261
wow this looks very impressive, bravo!
v11sman
I learned a lot of interesting things