03 / THE PERFORMANCE LAB

Performance.
With proof.

Explore repository-reported CPU, memory, and RPC observations alongside their workload, conditions, and limits.

xOneRS / SIGNER RECOVERYMerged · PR #319

52%

less replay CPU

A libsecp256k1 change reduced reported CPU work in a 20,000-block net2040 replay. Zero mismatches were reported in each run.

Read the original result

PROCESS CPU TIMELOWER IS BETTER

Before≈873 s

After≈417 s

Before: 872.8–874.7 CPU seconds. After: 416.0–417.3. Rounded bars show the repository-reported comparison.

Context matters. Shared Intel i5-13500 build host under heavy, variable load. This is CPU time for a replay, not wall-clock latency, follower throughput, or a comparison against another client. Repository-reported results; not independently reproduced here.

Workload, build, and measurement method

WHAT WAS MEASURED

Sequential replay from genesis to block 20,000, release build, nice -n 19, checkpoint-every 0. Each run reports 200 archive hash checks and 11 V2 epoch switches.

IMPLEMENTATION & EVIDENCE

Before: b055158. After: 5839f94. Change merged as PR #319; reported measurements were not independently reproduced here.

Open the linked engineering work

01 /

A result needs a workload

The existing experiments are self-comparisons or small samples. They do not establish a cross-client TPS ranking or universal hardware requirement.

Pin the client commits, compiler, hardware, and operating system.
Use the same network, block range, initial state, and retention mode.
Separate warm and cold runs; report repeats, variance, and failures.
Include correctness, wall time, tail latency, memory, and disk growth.

Benchmark contract Download the evidence record

KEEP THE CURIOSITY GOING

Another way in.

The client family Developer handbook Field notes