Examples

Scope: Days AGO. Run commands from the repository root. Examples for the frozen legacy simulator are on the Legacy Days page.

TCP smoke fixture

configs/benchmarks/tcp/fattree_k4_tcp_cubic_f16_smoke.toml is a k=4 fat tree with 16 CUBIC flows, small enough for a quick check of every engine:

cargo run --release --bin days -- --engine scalar \
  configs/benchmarks/tcp/fattree_k4_tcp_cubic_f16_smoke.toml

cargo run --release --bin days -- --engine cpu --workers 2 \
  configs/benchmarks/tcp/fattree_k4_tcp_cubic_f16_smoke.toml

TCP corpus

configs/benchmarks/tcp/ also holds k=16 and k=32 fat trees with one Reno or CUBIC flow per host (1,024 and 8,192 flows). configs/benchmarks/tcp/tcp-corpus-manifest.md lists each fixture's shape. The byte-identity campaign runs all five fixtures on every available backend.

E5 wide TCP

E5 is a family of k=32 fat trees with 8,192 hosts and one byte-terminated 1 MiB TCP flow per host:

• configs/benchmarks/evaluation/e5_wide_k32_q200.toml (Reno, 200-packet queues)
• configs/benchmarks/evaluation/e5_wide_k32_q200_cubic.toml (CUBIC, 200-packet queues)
• configs/benchmarks/evaluation/e5_wide_k32_q256.toml (Reno, 256-packet queues)

Run the Reno fixture on the CPU engine:

cargo run --release --bin days -- --engine cpu --workers 4 \
  configs/benchmarks/evaluation/e5_wide_k32_q200.toml

E6 constant-rate traffic

E6 is a family of k=32 fat trees with 8,192 constant-rate PacketDistribution flows at increasing offered load: configs/benchmarks/evaluation/e6_cbr_k32_load_01.toml, _10, _30, and _60.

cargo run --release --bin days -- --engine cpu --workers 4 \
  configs/benchmarks/evaluation/e6_cbr_k32_load_01.toml

Frontier fixture

configs/benchmarks/lookahead/rq9_frontier_closed_k32.toml defines 32 flow sets of 8,192 closed-loop Reno flows, 262,144 flows in total. A Scalar run:

cargo run --release --bin days -- --engine scalar \
  configs/benchmarks/lookahead/rq9_frontier_closed_k32.toml

On a device, the default channel-stream cap of 2,048 events can exceed device memory for this fixture; --channel-events-per-stream 256 lowers the starting cap. Capacity is refuse-or-run, so the cap changes memory use and retries, never the result.

Capacity warm start

A device run can save the capacities it converged on and pass them to a later run of the same fixture as first-attempt sizing:

cargo run --release --features metal --bin days -- --engine metal \
  configs/benchmarks/tcp/fattree_k4_tcp_cubic_f16_smoke.toml \
  --dump-capacity-warm-start /tmp/days-capacity-warm-start.txt

cargo run --release --features metal --bin days -- --engine metal \
  configs/benchmarks/tcp/fattree_k4_tcp_cubic_f16_smoke.toml \
  --capacity-warm-start /tmp/days-capacity-warm-start.txt

The snapshot changes initial arena sizing only. Both runs report the same result_bytes and result_fnv1a64. Use --features cuda and --engine cuda on an NVIDIA host.