Calibrate 3 tests for ADR-0033 Phase 2c per-flit wire timing
- test_h2d_local_cube_cut_through: threshold 65 → 80ns. The cut-through invariant (vs store-and-forward ~160ns at 4KB through UCIe) is what the test guards; the previous 65ns ceiling was too tight against the small per-flit overhead now charged at wire. - test_engine_override_is_scoped_to_impl: ZeroRouter inherits TransitComponent (was ComponentBase). Inheriting bare ComponentBase reverts the override path to non-flit-aware reassembly, making override slower than default and inverting the test. The test's intent is overhead=0 vs overhead=2, not flit-awareness. - test_intra_sip_critical_path_at_96k_below_threshold: threshold 20.5 → 30 µs. Allreduce absolute timing is sensitive to model fidelity; the algorithmic invariant (8-hop center root < 12-hop corner root) is preserved within the new envelope. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -143,10 +143,15 @@ def test_engine_override_is_scoped_to_impl():
|
||||
"""forwarding override (ZeroRouter, no overhead) reduces total_ns.
|
||||
|
||||
Router nodes have overhead_ns=2.0. Replacing with zero-latency impl
|
||||
removes router overhead from the path.
|
||||
removes router overhead from the path. The override class inherits
|
||||
from TransitComponent so it keeps flit-aware pass-through semantics
|
||||
(ADR-0033 Phase 2c); inheriting from bare ComponentBase would force
|
||||
per-hop flit reassembly = store-and-forward, making the override
|
||||
SLOWER than the default and inverting this test.
|
||||
"""
|
||||
from kernbench.components.builtin.forwarding import TransitComponent
|
||||
|
||||
class ZeroRouter(ComponentBase):
|
||||
class ZeroRouter(TransitComponent):
|
||||
def run(self, env, nbytes):
|
||||
yield env.timeout(0)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user