Performance optimization · Fleet result

# Compress the same data.   Get through it faster.

A fleet-produced Zstd change increased throughput by 7.81% on six public Silesia files, with identical compressed sizes.

The problem

## Compression runs  over and over.

In a data-intensive system, a small improvement to a heavily used operation can be worth pursuing. We asked a fleet to improve Zstd while preserving its compression settings and output sizes.

Six public Silesia files / independent replay 7.81% higher throughput
Than the pinned reference implementation. Identical compressed sizes; correct decoding.

## A change you  can build on.

The fleet returned a code patch that passed the original checks, upstream tests, boundary tests and cross-decoding checks. A separate replay measured the throughput improvement across the six files.

For your project, the same kind of assignment starts with the operation you want to improve, its workload and the behavior it must preserve.

Test setup, comparison and implementation status

The result is the geometric mean from nine paired measurements per file, using one-CPU-second timing windows and a pinned compiler, image and workstation. It is a public-corpus follow-up; the original private evaluation did not reach its 10% target.

A separate same-model solo run produced an eligible saved draft with a 0.65% gain against the same reference. It was selected retrospectively from four eligible drafts among eight submissions; no final draft was nominated. Both runs had roughly four active hours. The fleet grew from 16 to 1,024 agents with pauses, so this comparison does not hold population, spend or productive time constant.

This is an experimental implementation. A 32-bit portability concern remains open, and the change has not been accepted upstream. The result does not isolate the effect of agent communication.
[Result and measurement details ](/evidence/compression-result.json) [Zstd project ](https://github.com/facebook/zstd) 

## Put a fleet on your expensive operation.

256 agents. Four hours. $1,000 for your first run, with model inference, computers and fleet operation included.

[Explore your first fleet ](/first-fleet/) [Ask your agent what to run ](/evaluate/)