Add PCIe RX/TX load to the utilization charts

[Basavaraja-MS]

Summary

Closes #268.

Adds PCIe RX and TX load as plottable metrics in the line charts, alongside the existing GPU and memory utilization graphs.

• The load is derived from the already-collected pcie_rx / pcie_tx counters (KB/s) and normalized against the link's theoretical peak bandwidth (max_pcie_gen × max_pcie_link_width), so it fits the charts' fixed 0–100% scale.
• Two new metrics, pcieRxRate and pcieTxRate, are selectable per GPU from the F2 → Chart setup window and persist in the configuration file.
• The normalization math lives in a dedicated, unit-tested translation unit (src/pcie_utilization.c) rather than inline in interface.c.

Design note: max vs. current link

The metric is normalized against the maximum link bandwidth, not the currently negotiated link. GPUs down-train PCIe when idle (often to Gen1/x4), so normalizing against the live link would make tiny idle transfers look like high load and give a moving denominator. Using the max gives a stable "% of peak PCIe capability," matching the "spot the bottleneck" intent of #268. (The issue phrased it as "current link speed" — flagging this as a deliberate choice.)

Per-lane bandwidth basis

Effective per-lane data rates after PCIe encoding overhead (8b/10b for gen 1/2, 128b/130b for gen 3–5, FLIT for gen 6): 250 / 500 / 984.6 / 1969 / 3938 / 7563 MB/s.

Scope / limitations

• No layout or header changes — zero impact on users who don't enable the metrics.
• Only meaningful where the vendor backend reports both PCIe throughput and max_pcie_gen/max_pcie_link_width (NVIDIA, AMD, etc.); otherwise the line stays at 0, consistent with other unsupported metrics.
• Still bound by MAX_LINES_PER_PLOT (4 metrics per plot).

Test plan

• cmake -B build -DBUILD_TESTING=ON && cmake --build build — clean, no warnings.
• ctest — 8/8 pass, including 4 new PcieUtilization cases (per-gen bandwidth, width scaling, unknown-gen/missing-field → 0, percent scaling + clamp to 100%).
• Manual: ran nvbandwidth -t 0 -t 1 -b 256 -i 6 in a loop on an H200 NVL (Gen5 x16); enabled both metrics via F2. The GPU0 PCIe RX% / PCIe TX% lines peaked at ~95–99% under load and returned to 0 when idle, matching nvidia-smi dmon -s t (~60 GB/s of the 63 GB/s max).
• Config round-trips: pcieRxRate / pcieTxRate written to / read from the config file via F12.

Screenshot of the chart with the PCIe lines.

Made with Cursor