C++ 服务端 · 高性能系统 · 架构设计 · 工程效能 C++ Backend · High-Performance Systems · Architecture · Developer Productivity
专注于 C++ 服务端与高性能系统，习惯从架构设计、契约定义出发，把复杂系统拆成可设计、可验证、可交接的模块。对底层性能与工程化有执念：零依赖、可裁剪、缓存友好、可测试。
- 服务端 / 系统：高并发、资源调度、并发模型选型、可靠性与容错
- 底层 / 高性能：SIMD 向量化、内存池、缓存分块、访存优化
- 架构 / 工程：分层与边界、契约先行、开闭原则、构建与可观测性
语言
高性能 / 底层
服务端 / 工程
| 项目 | 简介 | 技术 |
|---|---|---|
| NetFountain | 两级代理 IP 池系统：一级池多供应商拉取净化去重，二级池按站点做出口连通验证形成租赁池，代理层按站点路由并原样透传，服务间以统一 HTTP API 进程级解耦 | Python · Vue 3 |
| PorxyPH | 代理节点编排管理系统：一份订阅拆为独立端口，每节点一容器、幂等生命周期、健康检查与自动故障转移 | Go · Vue 3 · Docker |
| FluxMeld | 本地多服务商 AI 路由工作台：OpenAI 兼容网关、账户池与健康感知负载均衡、安全工具调用 | TypeScript · Electron |
| SysTrace | 轻量级 Windows 系统资源历史监控：单文件部署，内置实时 CPU / 内存 / 磁盘热力图与可缩放时间线 | C++ |
| MemoryPool | 线程安全的高并发内存池，三层缓存架构，降低多线程下的锁竞争与分配开销 | C++ |
| MessageWall · Inbox · Cpp-Compiler · GTNH 核电 OC | 早期实践：校园信息墙服务、套接字通信、词法分析器、游戏内自动化 | C++ · Lua |
以下为内部 / 私有项目，仅作能力说明，不提供链接。
SLib — 自研神经网络算子库（手写 C++17）
- 零第三方依赖：C++17 + CMake 从零实现全部算子内核，不引入任何推理运行时或算子库，最大化可控性与可裁剪性。
- 高性能内核：手写 AVX2/FMA 微内核 SGEMM（寄存器分块、K 步长模板特化、软件预取、掩码加载）；向量化激活函数并以多项式逼近替换标准库数学函数（sin / exp / tanh / erf）。
- 卷积与流式状态：im2col + GEMM 通用卷积与因果卷积，设计滚动状态缓存支持流式推理；卷积分块参数按 L2 预算与 SIMD 宽度自动寻优。
- 信号处理：任意长度 FFT / STFT-ISTFT（2 的幂用迭代 FFT、任意长度用 Bluestein chirp-Z），含周期 Hann 窗、重叠相加与窗平方和归一化。
- 内存与抽象：对齐内存池
BufferPool+Buffer偏移视图，以零拷贝视图在算子链间传递张量；Module/Sequential统一算子接口，逐层复用消除反复分配。
服务端基础设施与高并发
- 资源调度与高可用：两级资源池、租约 + TTL、健康检查与自动故障转移。
- 并发治理：有界队列 + 固定 worker（队满丢最旧，长周期内存有界）、单写者模型、连接池串行 + keep-alive。
- 可靠性：按批推进 checkpoint、仅整批落盘后推进水位、写锁冲突保留队首重试、中断回滚未完成批。
- 容器编排：基于 Docker Engine API 的幂等容器生命周期（临时 init 容器写卷、按名清理）与周期健康检查。
数据工程
- 架构：共享核心库 + 薄壳 CLI，依赖单向、核心库零副作用。
- 扩展性：注册表 + 策略对象落地开闭原则，新增字段 / 规则 / 过滤策略不改既有流程。
- 数据管线：单写者串行化持久化、批量提交、流式写出与断点续传。
Focused on C++ backend and high-performance systems. I start from architecture, contracts, and module boundaries, and turn complex systems into units that are designable, verifiable, and handoff-ready. Strong bias toward zero dependencies, cache-friendly code, and testability.
- Backend / Systems: high concurrency, resource scheduling, concurrency-model selection, reliability and fault tolerance
- Low-level / Performance: SIMD vectorization, memory pooling, cache blocking, memory-access optimization
- Architecture / Engineering: layering and boundaries, contract-first design, open-closed principle, build and observability
Languages
High Performance / Low Level
Backend / Engineering
| Project | Description | Tech |
|---|---|---|
| NetFountain | Two-tier proxy IP pool: tier-1 pulls, sanitizes and deduplicates from multiple providers; tier-2 validates site-level connectivity into a leasing pool; a routing layer forwards per site, with services decoupled over a unified HTTP API | Python · Vue 3 |
| PorxyPH | Proxy node orchestration: one subscription split into standalone ports, one container per node, idempotent lifecycle, health checks and automatic failover | Go · Vue 3 · Docker |
| FluxMeld | Local multi-provider AI routing workspace: OpenAI-compatible gateway, account pool with health-aware load balancing, and safe tool calling | TypeScript · Electron |
| SysTrace | Lightweight Windows system-resource history monitor: single-file deployment with a built-in real-time CPU / memory / disk heatmap and zoomable timeline | C++ |
| MemoryPool | A thread-safe high-concurrency memory pool with a three-tier cache architecture to cut lock contention and allocation overhead | C++ |
| MessageWall · Inbox · Cpp-Compiler · GTNH Reactor | Early work: campus message wall, socket communication, a lexer, and in-game automation | C++ · Lua |
Internal / private projects, described only — no links.
SLib — a hand-written C++17 neural-network operator library
- Zero third-party dependencies: every operator kernel implemented from scratch with C++17 + CMake, no inference runtime or external operator library — maximizing control and trimmability.
- High-performance kernels: hand-written AVX2/FMA SGEMM micro-kernel (register blocking, K-step template specialization, software prefetch, masked loads); vectorized activations with polynomial approximations replacing libm math functions (sin / exp / tanh / erf).
- Convolution & streaming state: im2col + GEMM general and causal convolution with a rolling state cache for streaming inference; blocking parameters auto-tuned against L2 budget and SIMD width.
- Signal processing: arbitrary-length FFT / STFT-ISTFT (iterative FFT for powers of two, Bluestein chirp-Z otherwise) with periodic Hann window, overlap-add and window-sum normalization.
- Memory & abstraction: aligned memory pool
BufferPoolplusBufferoffset views, passing tensors across the operator chain as zero-copy views; unifiedModule/Sequentialinterfaces reuse buffers layer by layer.
Backend infrastructure & high concurrency
- Resource scheduling & HA: two-tier resource pools, lease + TTL, health checks and automatic failover.
- Concurrency governance: bounded queue + fixed workers (drop-oldest on full, bounded long-run memory), single-writer model, serial connection pool + keep-alive.
- Reliability: batch-wise checkpoint advance, commit-then-advance watermark, keep-head retry on write-lock contention, rollback of incomplete batches on interrupt.
- Container orchestration: idempotent container lifecycle via the Docker Engine API (ephemeral init container writing volumes, cleanup by name) with periodic health checks.
Data engineering
- Architecture: shared core library + thin CLI shells, unidirectional dependencies, side-effect-free core.
- Extensibility: registry + strategy objects implementing the open-closed principle so new fields, rules and filters require no changes to existing flows.
- Pipelines: single-writer serialized persistence, batched commits, streaming output and resumable execution.