Deployment Runbook

Deployment runbook

Shared CAPI/PSLSE environment setup, timeout controls, and generic host/RTL failure meanings are maintained in the CAPI-Precis deployment runbook. This page contains only AccelGraph-specific deployment evidence.

Before launch

git submodule sync --recursive
git submodule update --init --recursive
sudo apt-get install libjudy-dev verilator
verilator --version  # must report version 5 or newer
make verify

Record the graph file, format, algorithm, direction, CU count, FPGA image, thread counts, and timeout overrides.

Launch

The root wrapper supplies AccelGraph's graph-simulation directory and pinned CAPI-Precis checkout to the canonical environment harness:

./tools/capi-env --mode host check
./tools/capi-env --mode sim check
./tools/capi-env --mode sim -- make run-vsim
./tools/capi-env --mode sim -- make run-pslse
./tools/capi-env --mode sim -- make run-capi-sim-verbose2
./tools/capi-env --mode fpga -- make run-capi-fpga-verbose2

Use ./tools/capi-env --mode sim shell for a temporary configured shell. Exit the shell to discard the environment. The wrapper never edits or relies on .bashrc; mode behavior, custom Intel FPGA paths, and printed exports are defined by the CAPI-Precis environment harness. In simulation mode, CAPI_DEVICE must select the AFU configured in 03_capi_integration/accelerator_sim/server/shim_host.dat.

Simulation targets default to 5-minute start/call, 15-minute stall, and 2-hour run limits. Override the SIM_ACCELERATOR_* make variables when measured ModelSim latency requires different bounds.

Start with a TEST graph before moving to GAP, SNAP, LAW, or another large suite.

Failure evidence

Preserve the Accelerator verification failed line together with the command line, graph metadata, CU image, AFU/CU status, error register, progress counters, target vertex count, and simulator debug.log when applicable.

Reason	First check
device-error	WED pointers, graph size, MMIO error decode, selected image
stalled	Counter movement, frontier/iteration size, PSL response statistics
timeout	Workload scale, algorithm/direction/image match, measured healthy duration
MMIO failure	Device node, permissions, attach state, PSLSE health

Do not classify an algorithm as correct from liveness alone; compare its result against the existing reference path.