Blog
Blog
Case studies and measured findings. Each one traces to published raw artifacts, so any figure here can be checked against the data behind it. New posts are in the Atom feed.
ReleaseReamer Research 4.4.0 and 4.5.0: Larger-Than-RAM Data, Monte Carlo, Full Position State and a Relay to Reamer ServerReamer Research 4.4.0 runs datasets larger than RAM from memory-mapped bar files, adds Monte Carlo resampling and an equity curve, and gives every function a plain name. 4.5.0 gives a strategy its full position state and every resting order on each bar, lets it re-price orders, move stops and targets, close all or cancel all, adds a progress callback that can stop a run, and sends the same orders a backtest produces to a running Reamer Server. C ABI version 7, additive from here on.FindingsWhat 2,000 Runs on a 64-Core EPYC Show About Reamer ResearchA single-threaded research engine has one job the rest of research depends on: run the same input twice and get the same answer, at a speed that does not wander. I measured both across 2,000 runs on an AMD EPYC 9575F. Timing held to a 0.59% coefficient of variation and every hashed output came back byte-identical.FindingsReamer Server on a 64-Core EPYC: The Tail Is the MeasurementA trading engine is usually sold on one number, a peak throughput or a best-case latency. Neither tells a team what happens under concurrent load. I measured the integration boundary, the end-to-end order path, and the full 1-to-512 strategy sweep on an AMD EPYC 9575F. The boundary costs 53 nanoseconds, the engine-only P99 sits at 12.4 microseconds, and throughput holds a plateau to about 67 strategies.
Buy a licence30-day trial or annual licence. Key and kit by email on payment.