build: optimized-build mode + ns-3 submodule bump (Draft, blocked by aliyun/ns-3-alibabacloud#21)

[tianhao909]

PR-γ: default NS-3 optimized build mode + bump ns-3-alibabacloud submodule

Target repository: aliyun/SimAI

---

Summary

Two minimal build-system changes to aliyun/SimAI, meant to be reviewed together:

1. Flip the default NS-3 build mode from debug to optimized in astra-sim-alibabacloud/build/astra_ns3/build.sh (1-line change). The optimized profile roughly halves NS-3 simulation wall-clock time on the H20 microbenchmark set.
2. Bump the ns-3-alibabacloud submodule pointer from 7e3cb5b to the new upstream master HEAD produced by the companion PR-α (GCC 13 build fix + UB fix). The new hash is filled in at the moment PR-α is merged.

Together these let an out-of-the-box GCC 13 user run ./scripts/build.sh -c ns3 on aliyun/SimAI:master and get a working SimAI_simulator binary.

Key Changes

1. astra-sim-alibabacloud/build/astra_ns3/build.sh — 1 line changed
   • ./ns3 configure -d debug → ./ns3 configure -d optimized
2. Submodule pointer for ns-3-alibabacloud — 1 line changed (filled in after PR-α merges)
   • Before: 7e3cb5b88c99abcb582c5abc3919484a4805111b
   • After: (populated at push time)

Total diff: 2 lines across 2 files, 2 commits (build-mode commit, then bump commit).

Testing

Fingerprint (inline):

gcc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
Linux kernel 5.10.134-16.3.al8.x86_64 x86_64 GNU/Linux
Python 3.13.11

Baseline (upstream aliyun/SimAI:master @ f5efb5a + submodule at 7e3cb5b, GCC 13):

./scripts/build.sh -c ns3
# Expected: build FAILS with
#   bit-serializer.h:NN:NN: error: 'uint8_t' does not name a type
# This motivates the bump in this PR.

With this PR's HEAD (submodule pointing to PR-α merged HEAD, build.sh using -d optimized):

./scripts/build.sh -c ns3
file bin/SimAI_simulator
# Expected: build rc=0
# Expected: ELF 64-bit LSB executable ... not stripped
# The top-level scripts/build.sh rm -rf's extern/ and re-copies from
# the (now-patched) submodule on every clean build, so the GCC 13 fix
# is picked up automatically once the submodule pointer lands upstream.

End-to-end cross-branch smoke (all three PRs merged into a temporary integration branch, plan07 allreduce 16m case):

• RUN_EXIT=0
• SimAI binary sha256 recorded inline at push time in pr_drafts/task4_gamma_local_bump.txt
• Note: before aliyun/ns-3-alibabacloud actually merges PR-α, the pre-push verification uses a local simulation of "submodule pointed at PR-α's local branch HEAD" — this is documented in the internal test log and clearly separated from the real submodule pointer that ships with the PR commit.

Known Limitations

c_build — GCC 13 / Ubuntu 24.04 only. The bump is motivated and verified specifically for gcc-13.3.0 / Ubuntu 24.04; older toolchains (gcc-9, gcc-11) were not re-run against the new submodule HEAD in this work. The new submodule commits are additive (an include and a compile-flag pair), so compatibility is expected, but a CI run on the older matrix would be welcome.

Behavioural note on optimized default. Switching the default from debug to optimized is a behavioural change for downstream consumers who rely on unstripped, -O0 builds for in-process gdb sessions. -O2 -g preserves gdb attach-ability but does change the default numeric behaviour of stack frames. If maintainers prefer, this PR can be split into:

• γ1 — build.sh default mode only (1 line, no submodule change)
• γ2 — submodule bump only (1 line, no build.sh change)

The two commits are intentionally independent so either can be dropped.

Scope Disclaimer

English: This result does NOT serve as a proof of SimAI's generalization accuracy on H20/NVLink across all scenarios. It only reproduces a specific microbenchmark set under a specific calibration. Generalization to other models, sizes, topologies, or fused ops is not established.

中文：本结果不构成对 SimAI 在 H20/NVLink 全场景泛化准确性的证明。它仅复现特定 microbenchmark 集合在特定 calibration 下的精度。对其他模型、规模、拓扑、融合算子的泛化能力均未建立。

Notes

• Base: aliyun/SimAI:master (currently f5efb5a)
• Head: tianhao909/SimAI:pr/plan07-build-mode-and-bump
• Blocked by: aliyun/ns-3-alibabacloud#21 (PR-α, GCC 13 + UB fix). The submodule-bump commit is added only after PR-α merges, at which point this PR transitions from Draft to Ready.
• Independent of PR-β (topology/config), which can merge first.
• No rdma-hw.cc, no CMakeLists, no new runtime dependency.

Checklist

• Conventional commit messages (build: …)
• Diff restricted to build.sh + submodule pointer
• Scope Disclaimer inline (English + Chinese, > blockquote)
• c_build limitation called out in Known Limitations
• Clear "Blocked by PR-α" note in Notes
• Split-into-γ1/γ2 option documented if maintainers prefer single-topic PRs

Reviewer FAQ

Question	Answer
Why bump the submodule at all? Can I just use the latest aliyun/ns-3-alibabacloud HEAD?	The target HEAD is the one from PR-α, which contains the minimum patch set (4 files, +7 lines) verified against this repo's build. Newer ns-3 commits would pull in unreviewed behaviour.
Why change default to optimized?	It halves wall-clock time on the microbenchmark set this project actually runs. -g is preserved, so gdb still attaches. If preferred, I can leave the default at debug and add a SIMAI_BUILD_MODE env-var override instead — just say the word.
Can I keep the default debug?	Yes — split this PR into γ1 (build.sh only) and γ2 (submodule only), and drop γ1. See Known Limitations.
Does this affect astra-sim-alibabacloud/extern/ tracked files?	No. extern/ is .gitignore'd in this repo, and scripts/build.sh -c ns3 (lines 17 and 22) rm -rf extern/ then cp -r from the submodule. Nothing tracked under extern/ is preserved or affected.
Why not fix GCC 13 inside this repo directly?	The build actually compiles sources copied from the ns-3 submodule, so the fix has to live in aliyun/ns-3-alibabacloud. That's PR-α.
What happens if PR-α is rejected?	This PR cannot merge; the submodule-bump commit is withheld. γ1 (build-mode only) could still merge stand-alone if there is agreement on the optimized default.