Fair benchmark: random data, exclude dry run from metrics #10
Reference in New Issue
Block a user
Delete Branch "fair-benchmark"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Two fairness fixes for the benchmark suite:
Random data instead of all-zeros: Bulk files now use
random.randbytes()instead ofb"0" * n. Zero-filled data compresses to near-nothing, making zstd compression look unrealistically effective. Random data is incompressible, so speedups reflect real protocol efficiency.Dry run excluded from best-client metrics:
print_metricsnow filters out dry-run results, which transfer zero bytes but always appear as the "best" configuration. Real transfer configs are correctly compared against rsync.Before (zeros): compression won by compressing payload to near-zero
After (random): Standard (no metadata) is fastest at 0.81s, 6.79x vs rsync archive, 1.75x vs rsync compress — honest numbers from protocol architecture, not data artifacts.