[salimsuhet]

Problem

htsim_uec_df and htsim_uec_sf can't be run with different seeds, so repeated runs are not independent samples:

• Fixed global seed. Both binaries hardcode seed = 13 and, unlike htsim_uec, have no -seed option.
• Same switch RNG seed everywhere. Every DragonflySwitch and SlimFlySwitch seeds its std::mt19937 with the same value, 33. As a result, all switches draw the same random sequence. Under VALIANT/UGAL_L this means every switch picks the same intermediates in lockstep, which correlates path choices across the network.

Change (42 lines)

• New -seed N option in main_uec_df.cpp and main_uec_sf.cpp. It reseeds srand/srandom with N, which also affects the UEC entropy (EV) choices.
• Per-switch seeding. When -seed is given, each switch seeds its generator with std::seed_seq{N, switch_id}.
• Default unchanged. Without -seed, the binaries behave exactly as before (seed 13, and 33 for every switch).

Validation

Workload: 1024-host random permutation, 2 MB per flow, VALIANT routing (Dragonfly p4a8h4, SlimFly p7q9). The per-flow FCTs were compared across runs:

Test	Result
Patched binary without -seed vs. current main	identical (bit-for-bit)
-seed 1 vs. -seed 1	identical (still deterministic)
-seed 1 vs. -seed 2	different

Effect of per-switch seeding on the same workload:

Config	Mean FCT	Max FCT	Retransmissions
Dragonfly VALIANT, main (all switches seeded with 33)	91.6 µs	136.9 µs	6,482
Dragonfly VALIANT, -seed 1	90.1 µs	135.0 µs	3,901
Dragonfly VALIANT, -seed 2	89.9 µs	133.8 µs	4,003
SlimFly VALIANT, main	90.3 µs	126.6 µs	1,045
SlimFly VALIANT, -seed 1	90.5 µs	127.6 µs	1,000

FCT changes by only about 2%. On Dragonfly, however, retransmissions drop by about 40%. This is consistent with the shared seed concentrating VALIANT detours on the same intermediate groups. On SlimFly, the effect is negligible.

[tommasobo]

Thanks! I think this makes a lot of sense :)