Header: mnvkd combines zero-overhead abstractions, locality-aware memory, and a coroutine system3×faster than Go’s — achieving sub-millisecond latency and300K that scales like serverless — requests/sec/core on commodity hardware.
Performance
Benchmarks show over 300,000 full HTTP requests per second per core on a commodity i7 processor—delivering consistent sub-millisecond median latency and deterministic memory usage.
This level of throughput and stability is achieved through mnvkd’s unique combination of zero-overhead abstractions, locality-aware design, and compiler-level optimization.
Ontological Objects & Locality of Reference
DataVec introduces theontological object model—a new approach that complements traditional lambda calculus by embedding locality directly into computation. Every framework layer in mnvkd is cache-aware and designed to exploit modern CPU and virtual-memory architecture. While unrestricted C remains supported, mnvkd’s selective,cache-aware encapsulation delivers extreme efficiency without limiting developer freedom. These zero-overhead abstractions expose high-level, familiar interfaces—threaded I/O, OTP-style actors, and coroutine scheduling—while maintaining C-level performance. At its core,mnvkd coroutine continuations are 3× faster than Go goroutines, and total throughput improves further due to application code locality—unencumbered by Go’s work-stealing scheduler or garbage collection. Smaller cache sizes only amplify mnvkd’s advantage, since the system is explicitly tuned forspatial efficiency.
Key design advantages include:
- Library data and runtime stateco-located within a single huge-page micro-process
- Compilers can leveragelambda calculus optimizations when low-level calls do not escape
- Link-Time Optimization (LTO) applies across libraries, extending locality benefits beyond single-binary builds typical of Go or Rust
The result is a synthesis of functional and spatial reasoning: Functional paradigms simplify code and object management, while ontological objects enablemore effective function optimization — producing software that is both elegant and fast.
mnvkd V irtual Sockets
In mnvkd,virtual sockets (vk_vectoring) replace traditional stream-based I/O withbidirectional iovec ring buffers for TX/RX data. These buffers automatically wrap and integrate directly with mnvkd’s scheduler and network poller—eliminating syscalls, copy overhead, and flushing complexity.
Throughdeductive polling sockets register and flush automatically at optimal points, achieving optimal throughput with minimal system interaction.
Although built at a lower level, these I/O interfaces areeasier to use than conventional streams — even Python’s file objects—allowing developers to build high-performance, socket-based servers with minimal code.
Just as SQLite serves as a toolkit for building embedded databases,mnvkd serves as aserver authoring toolkit, enabling efficient, locality-aware network applications. In fact, the embedded nature of SQLite pairs naturally with mnvkd’slocality paradigm, creating a unified foundation for data-intensive, event-driven software.
