mirror of https://github.com/EasyTier/EasyTier.git synced 2026-08-06 12:39:51 +00:00

Files

T

fanyang 4b654fc56e perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

The async fn Future state machine overhead (~1.9us) dominated
MpscTunnelSender::send, while RingSink operations were only ~40ns.

Breakthrough: make send() an async fn that completes synchronously
on the first poll for the direct (ring tunnel) path. Uses
futures::task::noop_waker() to construct a dummy Context, then calls
Sink trait methods (poll_ready, start_send, poll_flush) directly.
RingSink always returns Ready immediately, so the waker is never
invoked and the async fn completes without yielding.

Channel mode (TCP/UDP/WG tunnels) still uses async send_async()
with proper backpressure. Ring tunnels detected via tunnel_info()
type check in PeerConn.

Results (4 threads, 1400B, 15s):
  pps:                   249K → 474K (+90%)
  send_msg_by_ip:        3.53us → 1.67us (-53%)
  send_msg_internal:     2.40us → 502ns  (-79%)
  MpscTunnelSender::send: 1.97us → 144ns  (-93%)

All 207 peers:: tests pass. Netns-requiring tests (three_node,
credential) unchanged (require root).

2026-06-28 21:15:05 +08:00

001-thread-safe-metrics-throughput.md

perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

2026-06-28 21:15:05 +08:00

002-bound-peer-rpc-queues.md

perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

2026-06-28 21:15:05 +08:00

003-avoid-noop-ospf-route-rebuilds.md

perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

2026-06-28 21:15:05 +08:00

004-reuse-ospf-route-graph.md

perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

2026-06-28 21:15:05 +08:00

005-cover-socks5-exit-node.md

perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

2026-06-28 21:15:05 +08:00

006-send-chain-cpu-optimization.md

perf(hotpath): add cpu_hotspot_ring bench and send-chain optimization plan

2026-06-28 12:45:05 +08:00

README.md

perf(mpsc): sync send via noop_waker — +90% pps (249K → 474K)

2026-06-28 21:15:05 +08:00

README.md

实施计划

由 improve skill 于 2026-06-18 生成，基于 commit 78146d16。除非依赖关系另有要求，请按以下顺序执行。每个执行者在开始前必须完整阅读计划，遵守 STOP 条件，并在完成后更新自己的状态行。

执行顺序与状态

Plan	标题	优先级	工作量	依赖	Status
001	将共享 metrics/throughput 计数改为线程安全实现	P1	M	—	DONE in worktree, not merged
002	为 peer RPC/control packet 队列加入背压和过载行为	P1	M	001	DONE in worktree, not merged
003	避免 OSPF 对 stale/no-op sync payload 重算路由	P2	S	—	DONE in worktree, not merged
004	复用 OSPF route-table 构图以减少拓扑更新成本	P2	M	003	DONE in worktree, not merged
005	补齐 SOCKS5 exit-node 集成测试覆盖	P2	M	—	DONE in worktree, not merged

Reconcile 2026-06-18

001: /home/fanmi/tmp/easytier-exec-001, branch advisor/001-thread-safe-metrics-throughput, commit 7b6e4dfe; worktree clean; not contained in main at 78146d16.
002: /home/fanmi/tmp/easytier-exec-002, branch advisor/002-bound-peer-rpc-queues, commit 34d2193d; worktree clean; not contained in main at 78146d16.
003: /home/fanmi/tmp/easytier-exec-003, branch advisor/003-avoid-noop-ospf-route-rebuilds, commit 1be77b51; worktree clean; not contained in main at 78146d16.
004: /home/fanmi/tmp/easytier-exec-004, branch advisor/004-reuse-ospf-route-graph, commit 325c2e5d; worktree clean; not contained in main at 78146d16.
005: /home/fanmi/tmp/easytier-exec-005, branch advisor/005-cover-socks5-exit-node, commit 2cb51b71; worktree clean; not contained in main at 78146d16.

依赖说明

002 依赖 001，因为队列背压计划应暴露 queue depth/drop counters；这些 counters 应复用 001 中线程安全后的 metrics primitive，避免在新代码里继续扩散 UnsafeCell 模式。
004 依赖 003，因为先抑制 no-op sync 的无效重算，再做共享构图重构，能让性能测试和行为变化更容易归因。
005 独立执行，但如果未来要修改 SOCKS5、exit-node 或 0.0.0.0/0 proxy CIDR 行为，应先落地 005 作为 characterization baseline。

已考虑并拒绝的发现

/api/v1/generate-config 和 /api/v1/parse-config 是否应要求登录：证据显示 route layering 可能使其公开，但可能是产品意图；不属于本次“正确性和性能”范围，且应先补意图测试再判断。
OSPF 7k 行 god module 整体拆分：确认是技术债，但范围过大；应先执行 003、004 并增加 characterization tests 后再规划。
前端测试/DX、依赖清理、安全 hardening：有价值，但用户本次只要求正确性和性能计划。