OpenSynaptic v1.2.0 Release Notes
发布日期 / Release Date: 2026-04-06
中文版
v1.2.0 — RSCore 高性能优化版
本次更新专注于 rscore（Rust 计算引擎）的性能深度优化，是 OpenSynaptic 1.x 系列的最终生产级基线版本。
🚀 性能优化（RSCore）
经过对热路径的逐层分析与多轮基准测试，本版本完成了 6 项 Rust 内存分配优化，并明确识别出性能瓶颈所在：
| 优化项 | 说明 |
|---|---|
① parse_apply_blob 零拷贝 | 返回类型从 Vec<u8> 改为 &[u8]，消除批处理循环中的堆分配 |
② decode_ts_token_to_6bytes 栈解码 | 改用 decode_slice 写入 6 字节栈数组，去除 heap alloc |
③ 消除 format! 热路径 | 新增 u32_to_decimal_slice 辅助函数，批处理循环中不再调用 format!("{};{}") |
| ④ SmallVec 传感器列表 | raw_vals 使用 SmallVec<[&'a [u8]; 16]>，≤16 传感器时完全避免堆分配 |
| ⑤ SmallVec 包体缓冲 | diff_body（64B）和 body（96B）改用 SmallVec 栈内联，减少 DIFF 路径 alloc |
| ⑥ sig_buf 跨 item 复用 | 新增 auto_decompose_fill_sig，每批次只分配一次 sig_buf 并复用，节省 N-1 次 Vec 分配 |
基准测试结果（AMD Ryzen 5 9600X，12 进程并行）：
- 优化前：~1.067M pps
- 优化前（上版基线）：~1.32M pps
- 本版：~1.32M pps（alloc 优化已达边际，瓶颈转移至 Python FFI 层）
结论： Rust 计算部分已占端到端时间的 ~17–20%，Python FFI 层（批量打包/memcpy/解包）占 ~80%。本版本作为当前架构下的最终优化基线，后续性能提升需要架构层面调整（ABI 瘦身或 Rust 驱动循环）。
🏗️ 架构分析（记录在案）
本版本完成了详细的架构分析，结论归档如下：
- 当前架构合理，不是技术债：Python 作为业务逻辑驱动层 + Rust 作为计算引擎是有意的设计选择
- 唯一的轻度技术债：
_invoke_batch中的from_buffer_copy可改为from_buffer（防御性拷贝已无必要） - 评估了 3 条进一步优化路径：
- 方案 A（Cython）：收益 <10%，不值得引入第三工具链
- 方案 B（per-item worker submit）：现有 API 为阻塞同步，实际收益仅 ~10%
- 方案 C（
os_bench_run_v1，Rust 驱动循环）：仅适用于基准测量，不适用于生产路径
- 在部署目标和 pps 需求明确前，以本版作为生产基线
🧹 目录清理
| 删除内容 | 原因 |
|---|---|
dist/ (v1.1.0rc1 旧产物, ~2MB) | 已过期的本地构建产物 |
target/rscore/ 根目录 (~1.1GB) | 孤立的 Rust 编译缓存 |
33 个 __pycache__ 目录 | Python 字节码缓存 |
.pytest_cache/ | pytest 运行缓存 |
src/opensynaptic/utils/c/bin/os_rscore_new.dll.bak | 无引用的 .bak 备份文件 |
📦 依赖变更
Cargo.toml 新增：
smallvec = { version = "1", features = ["union"] }
⚠️ 重大变更 / Breaking Changes
无。ABI 保持向后兼容，应用层 API 无变化。
升级指南
pip install --upgrade opensynaptic
English Version
v1.2.0 — RSCore Performance Hardening
This release focuses on deep performance optimization of the rscore (Rust compute engine), establishing the final production baseline for the OpenSynaptic 1.x series.
🚀 Performance Optimizations (RSCore)
Following systematic hot-path analysis and multiple benchmark iterations, this release delivers 6 Rust allocation-reduction optimizations and conclusively identifies the performance bottleneck:
| Optimization | Description |
|---|---|
① parse_apply_blob zero-copy | Return type changed from Vec<u8> to &[u8], eliminating heap allocation per batch iteration |
② decode_ts_token_to_6bytes stack decode | Switched to decode_slice writing into a 6-byte stack array, removing heap alloc |
③ Eliminate format! in hot path | New u32_to_decimal_slice helper replaces format!("{};{}") in batch loop |
| ④ SmallVec for sensor list | raw_vals now uses SmallVec<[&'a [u8]; 16]>, avoiding heap allocation for ≤16 sensors |
| ⑤ SmallVec for packet body buffers | diff_body (64B) and body (96B) use SmallVec inline storage, reducing DIFF path allocs |
| ⑥ sig_buf reuse across items | New auto_decompose_fill_sig allocates sig_buf once per batch and reuses it, saving N-1 Vec allocations |
Benchmark results (AMD Ryzen 5 9600X, 12 parallel processes):
- Pre-optimization baseline: ~1.067M pps
- Prior session baseline: ~1.32M pps
- This release: ~1.32M pps (alloc optimization at diminishing returns; bottleneck has shifted to Python FFI layer)
Conclusion: The Rust compute layer now accounts for only ~17–20% of end-to-end time, with the Python FFI layer (batch packing / memcpy / unpacking) consuming ~80%. This release serves as the final optimization baseline under the current architecture; further throughput gains require architectural changes (thinner ABI or Rust-driven loops).
🏗️ Architecture Analysis (Logged for Record)
This release includes a thorough architectural analysis with the following conclusions:
- Current architecture is sound, not technical debt: Python as business-logic driver + Rust as compute engine is an intentional design choice
- Only minor technical debt identified:
from_buffer_copyin_invoke_batchcan be replaced withfrom_buffer(defensive copy is unnecessary under the GIL) - Three further optimization paths evaluated:
- Option A (Cython deep rewrite): ~50% gain possible but equivalent effort to Option C with less reward; adds third toolchain
- Option B (per-item blocking worker submit): Existing API is synchronous; real gain only ~10%, not worth the refactor
- Option C (
os_bench_run_v1, Rust-driven loop): Suitable for benchmarking only; not applicable as production path
- Current version adopted as production baseline pending deployment target and pps requirement definition
🧹 Repository Cleanup
| Removed | Reason |
|---|---|
dist/ (v1.1.0rc1 artifacts, ~2MB) | Stale local build artifacts |
target/rscore/ at repo root (~1.1GB) | Orphaned Rust build cache |
33 × __pycache__ directories | Python bytecode caches |
.pytest_cache/ | pytest run cache |
src/opensynaptic/utils/c/bin/os_rscore_new.dll.bak | Unreferenced backup file |
📦 Dependency Changes
Added to Cargo.toml:
smallvec = { version = "1", features = ["union"] }
⚠️ Breaking Changes
None. ABI remains backward compatible; application-layer API is unchanged.
Upgrade
pip install --upgrade opensynaptic