Back to browse

Nexus – A 1KB .NET 10 core with <15μs latency for 1M parameters

by NexusCore·Mar 11, 2026·1 point·0 comments

AI Analysis

●●●●GemWizardryBig BrainNiche Gem

1KB DLL hitting 12.8μs on 1M entries with live telemetry proving it.

Strengths
  • Zero-allocation hot path eliminates GC pauses entirely for deterministic performance
  • Live telemetry dashboard streams real metrics via HTTP/3 QUIC for verification
  • 24-byte aligned structs optimized for CPU cache-line locality show genuine craft
Weaknesses
  • .NET 10 Native AOT requirement limits adoption to bleeding-edge stacks
  • 24MB memory-mapped file overhead may not suit memory-constrained deployments
Target Audience

.NET developers, performance engineers, systems programmers

Similar To

Chronicle · LMAX Disruptor · Agrona

Post Description

I’ve spent the last few weeks pushing the limits of .NET 10 (Native AOT) to see if I could achieve deterministic, hard real-time performance for large-scale parameter indexing. Today, I hit the 1,000,000 entry milestone with a search latency of ~12.8μs (0.0128ms). The Architecture (The NANO Standard): Zero-Allocation: The hot path is entirely heap-free. No GC pauses, no jitter. Memory-Mapped Index: I’m using a pre-allocated 24MB MMF acting as a "Software-Defined Hardware" surface. Binary Search: O(log n) lookups on 24-byte aligned structs, optimized for CPU cache-line locality. 1KB Footprint: The core logic is compiled down to a 1KB Native AOT DLL. The Telemetry Bridge: To verify the results, I built a decoupled telemetry pipeline. A local Avalonia dashboard reads the MMF via raw pointers and streams the metrics via HTTP/3 (QUIC) to a Next.js/Vercel Edge dashboard. Live Demo: You can see the live trace of the 1,000,000 entry index here: telemetry.intelligentaudio.net I’m happy to answer any questions about the pointer-arithmetic, the MMF-layout, or how I’m bypassing the managed heap entirely.

Similar Projects