What problem does it solve?
Deploying a custom Programmable Congestion Control algorithm on a live RDMA / RoCE fabric is high-stakes: a wrong or unstable algorithm can cause flow collapse, queue blow-up, or fabric-wide instability, and an idle-link evaluation produces no meaningful signal. This Skill gives agents the operator-side discipline for the doca_spcx_cc CLI so SPCX algorithms are loaded, evaluated under real contention, and rolled forward only with evidence and a rehearsed rollback.
Core Features & Use Cases
- Surface selection decision tree: Routes between SPCX, the established
doca-pcc surface, and factory firmware PCC before any code is written, based on install, BlueField generation, and algorithm needs.
- Contention-positive evaluation workflow: Walks the prepare → smoke → contention-injected run flow, pairing tracer output with PCC counter snapshots and RDMA-side metrics as evidence.
- Layered error taxonomy and safety gates: Diagnoses failures across install, device-binding, firmware-slot, DPA-image, algorithm-precondition, live-link, runtime, and version layers, and gates production cutover on bounded blast radius, OOB access, and rehearsed factory-PCC rollback.
- Use Case: A developer has DPACC-compiled a custom RTT-based congestion-control algorithm and a non-prod BlueField pair; the Skill guides loading it via
doca_spcx_cc, injecting matching contention, capturing the evidence tuple, and deciding whether it is safe to roll forward.
Quick Start
Ask your agent to load and evaluate your DPACC-compiled SPCX congestion-control algorithm on a non-production BlueField pair using doca_spcx_cc, with contention injected and a documented rollback to factory PCC.