Repository navigation
feat: performance optimization phase 3 — zero-copy frame pool, per-worker SPSC - #25
Merged
Merged
Conversation
…rker SPSC, RSS affinity Multi-core pipeline redesign addressing the remaining performance gaps: - P3.1: FramePool slab allocator — pre-allocated contiguous buffer of N×frame_size bytes with lock-free SPSC free list. Eliminates all per-packet heap allocation on the RX→Worker hot path. - P3.2-P3.3: FrameRef (pool_idx + len) replaces Vec<u8> in worker SPSC rings. Frames are passed by 8-byte reference instead of cloned Vec, zero-copy after initial DMA. - P3.4: Per-worker SPSC app rings replace shared MPSC app_ring. recv_from() polls round-robin across worker rings, eliminating CAS contention that scaled poorly with worker count. - P3.6: RSS-aware worker affinity — direct queue-to-worker mapping preserves flow locality from NIC RSS hashing. - P3.8: Comprehensive unit tests for FramePool (alloc/free cycles, pool exhaustion, producer-consumer with SPSC ring, concurrent access). https://claude.ai/code/session_01AX56ktjxp4ybrKFqAg7mMu
Mark P3.1-P3.4, P3.6, P3.8 complete. Add preliminary Run #3 entry to perf-test-log.md with pending benchmark results (perf-tests.yml running). https://claude.ai/code/session_01AX56ktjxp4ybrKFqAg7mMu
[Perf] Stage: DeployDeploying |
[Perf] Stage: Instances Ready
|
[Perf] Stage: TRex ConfigStarting TRex configuration (MAC discovery + NIC binding)... |
[Perf] Stage: TRex Config OK
|
[Perf] Stage: TRex StartedTRex server running. Beginning benchmarks... |
[Perf] DUT ReadyDUT instance |
[Perf] Stage: Benchmark (1/5)Running |
[Perf] Benchmark Diag:
|
[CI] Stage: DeployInfrastructure ready.
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (2/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (3/5)Running |
[Perf] Benchmark Diag:
|
[CI] Stage: SummaryAll tests PASSED. ARP seeding: kernel /proc/net/arp (automatic)
|
✅ Integration Tests Passed (Run 23035234544)Branch: Test Results
Application Logs (last 20 lines)receiver-echo-server.log sender-echo-server.log sender-test-client.log receiver-test-client-iperf.log sender-test-client-iperf.log Full Application Logs (last 200 lines each)receiver-echo-server.logsender-echo-server.logsender-test-client.logreceiver-test-client-iperf.logsender-test-client-iperf.log
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (4/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (5/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Diag: testpmd logtestpmd output (last 30 lines) |
[Perf] Stage: Results[04:39:58] INFO Generating markdown summary... Performance Test Results — c5n.2xlargeCommit: 1400B packets
512B packets
64B packets
|
…rings Two critical fixes for Phase 3 performance: 1. FramePool::free() race condition: Changed from load+store to fetch_add for MPSC safety when multiple workers free concurrently. Increased free list to 2x capacity to prevent wraparound collisions. 2. True zero-copy: Workers now create AppPacket (FrameRef + payload offset) instead of ProcessedPacket (Vec<u8>). Frames stay in pool until recv_from() copies payload to user buffer and frees. Eliminates the per-packet heap allocation that negated FramePool's benefit. https://claude.ai/code/session_01AX56ktjxp4ybrKFqAg7mMu
Owner
Author
Phase 3 Fix: True Zero-Copy PipelineThe initial Phase 3 benchmark showed a regression because:
Changes
Re-running perf tests: https://github.com/gspivey/dpdk-stdlib-rust/actions/runs/23036730290 |
[Perf] Stage: TRex Config OK
|
[Perf] Stage: TRex StartedTRex server running. Beginning benchmarks... |
[Perf] DUT ReadyDUT instance |
[Perf] Stage: Benchmark (1/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (2/5)Running |
[Perf] Benchmark Diag:
|
[CI] Stage: SummaryAll tests PASSED. ARP seeding: kernel /proc/net/arp (automatic)
|
✅ Integration Tests Passed (Run 23077712886)Branch: Test Results
Application Logs (last 20 lines)receiver-echo-server.log sender-echo-server.log sender-test-client.log receiver-test-client-iperf.log sender-test-client-iperf.log Full Application Logs (last 200 lines each)receiver-echo-server.logsender-echo-server.logsender-test-client.logreceiver-test-client-iperf.logsender-test-client-iperf.log
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (3/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (4/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Stage: Benchmark (5/5)Running |
[Perf] Benchmark Diag:
|
[Perf] Benchmark Diag:
|
[Perf] Diag: testpmd logtestpmd output (last 30 lines) |
[Perf] Stage: Results[02:47:41] INFO Generating markdown summary... Performance Test Results — c5n.2xlargeCommit: 1400B packets
512B packets
64B packets
|
gspivey
pushed a commit
that referenced
this pull request
May 21, 2026
…dmap Performance test results from GH Actions run 26204734648 show no regression from the IPv6 outer encap feature: - rust-dpdk 64B/700K: 699,000 RX (0.1% drop) — identical to Run #18 - rust-dpdk 512B/700K: 699,000 RX (0.1% drop) — identical to Run #18 - NIC instrumentation self-check: OK (zero drift) README roadmap updated to move 'Encap: IPv6 outer' from Planned to Done.
gspivey
pushed a commit
that referenced
this pull request
May 21, 2026
Append performance test results from GH Actions run 26227356354 (Graviton, TRex). No regressions detected — IPv6 fallback path adds zero overhead to IPv4 traffic. Mark IPv6 sub-task #2 (UDP over IPv6 checksum) as complete in the README roadmap.
gspivey
pushed a commit
that referenced
this pull request
May 29, 2026
Performance test results from GH Actions run 26633424088: - rust-dpdk 700K/64B: 695,587 RX (0.6% drop) — no regression vs Run #25 - rust-dpdk 700K/512B: 693,903 RX (0.9% drop) — no regression - tokio-dpdk caps at ~311K PPS — unchanged - native-dpdk baseline: 698,590 at 700K/64B Conclusion: IPv6 socket address support is performance-neutral.
gspivey
pushed a commit
that referenced
this pull request
Jun 16, 2026
gspivey
added a commit
that referenced
this pull request
Jun 16, 2026
## Roadmap Item `ROADMAP.md` item #25: `dpdk-stdlib-tcp`: Engine — on_command Spec: `.kiro/specs/tcp-support/` · tasks 5.18, 5.19 ## Changes Implements `TcpEngine::on_command` for the remaining three command variants: ### Shutdown (task 5.19) - **Shutdown::Write / Both**: Flush-before-FIN semantics — drains tx_ring into send_buf immediately, sets `fin_pending` flag. The `on_tick` TX drain path transmits all remaining data then emits FIN once send_buf is empty. - **Shutdown::Read**: Sets EOF on ConnectionHandle so app-side reads return 0. No FIN sent. - State transitions: Established → FinWait1, CloseWait → LastAck. ### Close (task 5.19) - **linger = None or linger > 0**: Graceful FIN (same as Shutdown::Write). App-side blocking handled via condvar. - **linger = Some(Duration::ZERO)**: Send RST immediately, discard all unsent data, remove TCB, latch ConnectionAborted. ### SetOption (task 5.19) - Nodelay, Keepalive (enable/disable clears timer), Linger (updates both TCB and handle), RecvBufSize, SendBufSize, ReuseAddr, Ttl. - ReadTimeout/WriteTimeout/Nonblocking are app-side only (no engine action needed). ### Infrastructure - `send_fin()` helper: builds FIN+ACK frame, advances snd_nxt, transitions state, arms RTO. - `fin_pending` handling in `on_tick`: after TX drain completes with no unsent data remaining, emit FIN via `send_fin()`. ## Tests Added 28 new unit tests covering: - Shutdown with empty/pending send_buf, Read/Write/Both variants - FIN sequence number consumption and RTO arming - Close with default/non-zero/zero linger - RST on linger=0 discards pending data - SetOption for all engine-relevant variants - Flush-before-FIN integration (data segment + FIN in correct order) - Edge cases: nonexistent key, shutdown in SYN_SENT ## Test Results All 192 dpdk-stdlib-tcp tests pass. Full workspace builds and all tests pass. --------- Co-authored-by: Agent Router <agent@agent-router.dev>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Multi-core pipeline redesign (Phase 3) targeting the remaining performance gaps at 350K-700K PPS:
FrameRef(pool_idx + len) replacesVec<u8>in worker SPSC rings. Frames passed by reference, not cloned.app_ringwith per-worker SPSC rings.recv_from()polls round-robin, eliminating CAS contention.Expected Impact (based on Phase 2 baseline)
Phase 2 Baseline (to beat)
Test plan
https://claude.ai/code/session_01AX56ktjxp4ybrKFqAg7mMu