Symptom: Many-core system shows escalating latency after 16 cores active. Snoop rate is manageable. Retry rate spikes.
Profile retry rate per home node — is one HN overloaded?
Check RN-to-HN routing — are requests evenly distributed?
If the selected home implements a directory/filter, measure that lookup and transaction-entry occupancy; do not assume its presence from the CHI label.
Correlate Protocol Credits, link credits and any implementation-specific NoC virtual-channel pools separately; a VC is not a CHI Protocol Credit.
Change home placement or resources only after distinguishing admission pressure from downstream delay and verifying the bottleneck under the same workload.
Treating CHI as 'AXI with snoops' — different topology, different debug.
Ignoring retry counters in performance analysis.
Assuming protocol family alone determines snoop filtering, topology or performance.