FEC Observability Loop · Arista ⇄ Cumulus
prometheus scrape · normalization · CUDA inference · self-healing optics
A-side / Z-side telemetry → CUDA → self-healing feedback loop
10 Hz · dual-vendor fusionA-side · arista-spine-07
Arista EOS 4.32 · gNMI / OpenConfig · Ethernet17/1 · 800G-DR8
Z-side · cumulus-leaf-22
NVIDIA Cumulus Linux 5.9 · NetQ / gNMI · swp23s0 · 800G-DR8
01
Prometheus scrape
gNMI-gateway exporters
02
Normalizer
units · rates · PTP align
03
CUDA inference
A100 · FEC tensor kernels
04
Policy LUT
signature → action
05
Actuate optics
EOS + Cumulus APIs
06
Verify / feedback
post-FEC re-measure
pre-FEC BER (A)
—
limit 2.4e-4
pre-FEC BER (Z)
—
KP4 RS(544,514)
post-FEC BER
—
SLO < 1.0e-12
corrected bits/s
—
FEC workload A-side
eSNR margin
—
Rx — dBm
closed loops
0
CUDA TTR ~1.9 s
Telemetry normalization · vendor path → Luma canonical metric
| luma canonical | arista eos (gNMI) | cumulus linux (NetQ/gNMI) | unit |
|---|---|---|---|
| link.pre_fec_ber | /interfaces/interface/state/counters/fec-uncorrectable… | netq · ethtool -S rs_fec_corrected | ratio |
| link.post_fec_ber | openconfig-platform-transceiver:post-fec-ber | sysfs · fec_uncorrected_blocks | ratio |
| optic.rx_power_dbm | …/physical-channel/state/input-power/instant | smonctl · sfp rx_power (mW) | dBm |
| optic.temp_c | …/transceiver/state/temperature/instant | hwmon · temp1_input (m°C) | °C |
| optic.tx_bias_ma | …/physical-channel/state/laser-bias-current | eeprom · tx_bias (µA) | mA |
| link.esnr_db | eos-native · phy snr per lane | mlxlink --show_eye · snr | dB |
| link.flaps_per_min | …/interface/state/counters/carrier-transitions | switchd · link_flap_count | 1/min |
Both endpoints are resampled to a common 10 Hz grid, unit-converted (mW→dBm, counters→rates), and aligned to PTP timestamps before the A/Z pair is fused into a single link object.
Execution log
awaiting events…