spargio

spargio is a work-stealing io_uring-based async runtime for Rust, using msg_ring for cross-thread coordination.

Instead of a strict thread-per-core/share-nothing execution model like other io_uring runtimes (glommio/monoio/compio and tokio_uring), spargio uses submission-time steering of stealable tasks across threads (a novel form of work-stealing).

In our benchmarks (detailed below), spargio outperforms compio (and likely all share-nothing runtimes) in imbalanced workloads by up to 70%, and outperforms tokio for cases involving high coordination or disk I/O by up to 320%. compio leads for sustained, balanced workloads by up to 70%.

Out-of-the-box, we support async disk I/O, network I/O (including TLS/WebSockets/QUIC), process execution, and signal handling, and provide an extension API for additional io_uring operations. We support both tokio-style stealable tasks and compio-style pinned (thread-affine) tasks.

Disclaimer

spargio began as an experimental proof-of-concept built with Codex. I have not manually reviewed all the code yet. Use for evaluation only.

Quick start

Pre-requisites: Linux 6.0+ recommended (5.18+ for core io_uring + msg_ring paths)

Add spargio as a dependency:

cargo add spargio --features macros,uring-native

Then use it for native I/O operations and stealable task spawning:

use spargio::{fs::File, net::TcpListener, RuntimeHandle};

#[spargio::main]
async fn main(handle: RuntimeHandle) -> std::io::Result<()> {
    std::fs::create_dir_all("ingest-out")?;
    let listener = TcpListener::bind(handle.clone(), "127.0.0.1:7001").await?;
    let mut id = 0u64;

    loop {
        let (stream, _) = listener.accept_round_robin().await?;
        let (h, s, path) = (handle.clone(), stream.clone(), format!("ingest-out/{id}.bin"));
        id += 1;

        stream.spawn_stealable_on_session(&handle, async move {
            let file = File::create(h, path).await.unwrap();
            let (n, buf) = s.recv_owned(vec![0; 64 * 1024]).await.unwrap();
            file.write_all_at(0, &buf[..n]).await.unwrap();
            file.fsync().await.unwrap();
        }).expect("spawn");
    }
}

Tokio Integration

Recommended model today:

• Run Tokio and Spargio side-by-side.
• Exchange work/results through explicit boundaries (spargio::boundary, channels, adapters).
• Move selected hot paths into Spargio without forcing full dependency migration.

Note: uniquely to Spargio, a Tokio-compat readiness shim based on IORING_OP_POLL_ADD is possible to build on top of it without sacrificing work-stealing, but building and maintaining a dependency-transparent drop-in lane would be a large investment.

Inspirations and Further Reading

Using msg_ring for coordination is heavily inspired by ourio. We extend that idea to work-stealing.

Wondering whether to build a work-stealing pool using io_uring at all was inspired by the following (excellent) blog posts:

• https://emschwartz.me/async-rust-can-be-a-pleasure-to-work-with-without-send-sync-static/
• https://without.boats/blog/thread-per-core/

Terminology: Shards

In Spargio, a shard is one worker thread + its io_uring ring (SQ + CQ) + a local run/command queue. Internally within Spargio, we pass work from one shard to another by enqueueing work and injecting CQEs across shards, waking up a recipient worker thread to drain pending work from its queue.

Benchmark Results

Coordination-focused workloads (Tokio vs Spargio)

Compio is not listed in this coordination-only table because it is share-nothing (thread-per-core), while these cases are focused on cross-shard coordination behavior.

Native API workloads (Tokio vs Spargio vs Compio)

Imbalanced Native API workloads (Tokio vs Spargio vs Compio)

Benchmark Interpretation

TL;DR: As expected, Spargio is strongest on coordination-heavy and low-depth latency workloads; Compio is strongest on sustained balanced stream throughput. Tokio is near parity with Spargio on rotating-hotspot network shapes.

• Spargio leads in coordination-heavy cross-shard cases versus Tokio (steady_ping_pong_rtt, steady_one_way_send_drain, cold_start_ping_pong, fanout_fanin_*).
• Spargio leads in low-depth fs/net latency (fs_read_rtt_4k, net_echo_rtt_256b) versus both Tokio and Compio.
• Compio leads in sustained balanced stream throughput and static-hotspot imbalance (net_stream_throughput_4k_window32, net_stream_imbalanced_4k_hot1_light7), while Spargio is currently ahead of Tokio in both of those cases.
• Tokio and Spargio are near parity in rotating-hotspot stream/pipeline cases and keyed routing (net_stream_hotspot_rotation_4k, net_pipeline_hotspot_rotation_4k_window32, net_keyed_hotspot_rotation_4k).

For performance, different workload shapes favor different runtimes.

What's Done

• Sharded runtime with Linux IoUring backend.
• Cross-shard typed/raw messaging, nowait sends, batching, and flush tickets.
• Placement APIs: Pinned, RoundRobin, Sticky, Stealable, StealablePreferred.
• Work-stealing scheduler with adaptive steal gating/backoff, victim probing, batch steals, wake coalescing, backpressure, and runtime stats.
• Runtime primitives: sleep, sleep_until, timeout, timeout_at, Interval/interval_at, Sleep (resettable deadline timer), CancellationToken, and TaskGroup cooperative cancellation.
• Runtime entry ergonomics: async-first spargio::run(...), spargio::run_with(builder, ...), and optional #[spargio::main(...)] via macros.
• Runtime utility bridge knobs: RuntimeHandle::spawn_blocking(...) and RuntimeBuilder::thread_affinity(...).
• Local !Send ergonomics: run_local_on(...) and RuntimeHandle::spawn_local_on(...) for shard-pinned local futures.
• Unbound native API: RuntimeHandle::uring_native_unbound() -> UringNativeAny with file ops (read_at, read_at_into, write_at, fsync) and stream/socket ops (recv, send, send_owned, recv_owned, send_all_batch, recv_multishot_segments), plus submission-time shard selector, FD affinity leases, and active op route tracking.
• Low-level unsafe native extension API: UringNativeAny::{submit_unsafe, submit_unsafe_on_shard} for custom SQE/CQE workflows in external extensions.
• Safe native extension wrapper slice + cookbook: spargio::extension::fs::{statx, statx_on_shard, statx_or_metadata} plus docs/native_extension_cookbook.md.
• Ergonomic fs/net APIs on top of native I/O: spargio::fs::{OpenOptions, File} plus path helpers (create_dir*, rename, remove_*, metadata/link helpers, read/write), and spargio::net::{TcpListener, TcpStream, UdpSocket, UnixListener, UnixStream, UnixDatagram}.
• Directory traversal + du parity helpers: low-level spargio::extension::fs::read_dir_entries(...) and high-level spargio::fs::{read_dir(...), du(...), DuOptions, DuSummary} with sparse/hardlink/symlink and one-filesystem policy support.
• Measured metadata fast path helper: spargio::fs::metadata_lite(...) (statx-backed with fallback).
• Native-first fs path-op lane on Linux io_uring for high-value helpers (create_dir, remove_file, remove_dir, rename, hard_link, symlink), with compatibility fallback on unsupported opcode kernels.
• Foundational I/O utility layer: spargio::io::{AsyncRead, AsyncWrite, split, copy_to_vec, BufReader, BufWriter} and io::framed::LengthDelimited.
• Native setup path on Linux io_uring lane: open/connect/accept are nonblocking and routed through native setup ops (no helper-thread run_blocking wrappers in public fs/net setup APIs).
• Native timeout path on io_uring lane: UringNativeAny::sleep(...) and shard-context spargio::sleep(...) route through IORING_OP_TIMEOUT.
• Async-first boundary APIs: call, call_with_timeout, recv, recv_timeout, and BoundaryTicket::wait_timeout.
• Explicit socket-address APIs that bypass DNS resolution: connect_socket_addr* and bind_socket_addr.
• Benchmark suites: benches/ping_pong.rs, benches/fanout_fanin.rs, benches/fs_api.rs (Tokio/Spargio/Compio), and benches/net_api.rs (Tokio/Spargio/Compio).
• Scheduler profiling lane with callgrind/cachegrind: scripts/bench_scheduler_profile.sh and ratio guardrail helper scripts/scheduler_profile_guardrail.sh.
• Mixed-runtime boundary API: spargio::boundary.
• Companion crate suite: spargio-process, spargio-signal, spargio-protocols (legacy blocking bridge helpers), spargio-tls (rustls/futures-rustls adapter), spargio-ws (async-tungstenite adapter), and spargio-quic with selectable backend mode (QuicBackend::Native default dispatch and explicit QuicBackend::Bridge compatibility fallback).
• Native-vs-bridge QUIC cutover guardrails: native data path is validated to avoid bridge task spawning, while bridge mode remains explicit compatibility fallback.
• QUIC native default backend now runs on quinn-proto driver path (NativeProtoDriver + native UDP pump/timers) with stream/datagram operations routed through the driver; bridge mode remains explicit compatibility fallback.
• Companion hardening lane: scripts/companion_ci_smoke.sh plus CI companion-matrix job.
• QUIC qualification lanes: interop matrix (scripts/quic_interop_matrix.sh), soak/fault lane (scripts/quic_soak_fault.sh, nightly), and native-vs-bridge perf gate (scripts/quic_perf_gate.sh).
• In-repo user-facing book/ (mdBook) covering quick start, task placement (!Send + stealable locality-first defaults), I/O API selection, protocol crates, native extensions, performance tuning, operations, migration, and status.
• Reference mixed-mode service example.

What's Not Done Yet

• Hostname-based ToSocketAddrs connect/bind paths can still block for DNS resolution; use explicit SocketAddr APIs (connect_socket_addr*, bind_socket_addr) for strictly non-DNS data-plane paths.
• Remaining fs helper migration to native io_uring where it is not a clear win is deferred: canonicalize, metadata, symlink_metadata, and set_permissions currently use compatibility blocking paths (create_dir_all is native-first for straightforward paths; metadata_lite exists as native-first metadata alternative).
• Work-stealing tuning guidance still needs deeper production case studies and calibration examples on top of the current knob and profiling documentation.
• Work-stealing queue structure is still generic compared with Tokio's specialized scheduler queues; evaluate a better-optimized queue design (owner-fast-path + injection/steal structure) later.
• Continue readability/editorial cleanup across README + book/: tighten wording, keep examples minimal but practical, and reduce ambiguous terminology.
• Broaden documentation coverage while refactoring core modules for maintainability: keep API docs/book content aligned as runtime/fs/net surfaces continue to be split into smaller focused units.

Longer-term Improvement Ideas

• Optional Tokio-compat readiness emulation shim (IORING_OP_POLL_ADD) is explicitly deprioritized for now (backlog-only, not planned right now).
• Full production-grade higher-level ecosystem parity is still in progress; companion crates now provide practical bridges and qualification lanes, but deeper protocol-specific maturity remains (broader TLS/WS tuning surfaces, richer process stdio orchestration, and deeper long-window failure coverage).
• QUIC backend hardening is still in progress: native default path is driver-backed now, but long-window soak/fault/perf requalification depth and rollout maturity (rollout_stage) still need production validation.
• Production hardening beyond smoke lanes: deeper failure-injection/soak coverage, broader observability for companion protocol paths, and long-window p95/p99 gates.
• Further workload-specific work-stealing model calibration is still iterative (the adaptive policy is implemented, but thresholds/weights are expected to continue evolving with production traces).
• Multi-endpoint QUIC sharding/fan-out orchestration is not built in yet: a single QuicEndpoint still owns one native transport backend, so multi-core listener scaling is currently a manual multi-endpoint deployment pattern.
• Fully io_uring-submitted directory traversal is still in progress: read_dir/du APIs are built-in, but (as of 2026-03-03) upstream io_uring userspace/kernel ABIs do not expose a stable getdents opcode surface (IORING_OP_GETDENTS), so traversal currently uses a blocking-helper lane (getdents64 with compatibility fallback) instead of pure in-ring submission.

Contributor Quick Start

cargo test
cargo test --features uring-native
cargo bench --features uring-native --no-run
cargo test --features macros --test entry_macro_tdd

Benchmark helpers:

./scripts/bench_fanout_smoke.sh
./scripts/bench_ping_guardrail.sh
./scripts/bench_fanout_guardrail.sh
./scripts/bench_kpi_guardrail.sh
./scripts/bench_scheduler_profile.sh
./scripts/bench_scheduler_calibration.sh
./scripts/scheduler_profile_guardrail.sh
./scripts/companion_ci_smoke.sh
./scripts/quic_interop_matrix.sh
./scripts/quic_perf_gate.sh
./scripts/quic_soak_fault.sh

Reference app:

cargo run --example mixed_mode_service

Runtime Entry

Helper-based entry:

#[tokio::main]
async fn main() -> Result<(), spargio::RuntimeError> {
    spargio::run(|handle| async move {
        let job = handle.spawn_stealable(async { 42usize }).expect("spawn");
        assert_eq!(job.await.expect("join"), 42);
    })
    .await
}

Attribute-macro entry (enable with --features macros):

#[spargio::main(shards = 4, backend = "io_uring")]
async fn main() {
    // async body runs on Spargio runtime
}

This takes two optional arguments. Without them, #[spargio::main] uses sensible defaults: io_uring backend and shard count from available CPU parallelism. Use macro arguments only when you need explicit overrides.

Repository Map

• src/lib.rs: runtime implementation.
• tests/: TDD coverage.
• benches/: Criterion benchmarks.
• examples/: mixed-mode reference app.
• scripts/: benchmark smoke/guard helpers.
• .github/workflows/: CI gates.
• IMPLEMENTATION_LOG.md: implementation and benchmark log.
• architecture_decision_records/: ADRs.

Connection Placement Best Practices

• Use spargio::net::TcpStream::connect(...) for simple or latency-first paths (few streams, short-lived connections).
• Use spargio::net::TcpStream::connect_many_round_robin(...) (or connect_with_session_policy(..., RoundRobin)) for sustained multi-stream throughput workloads.
• For per-stream hot I/O loops, pair round-robin stream setup with stream.spawn_on_session(...) to keep execution aligned with the stream session shard.
• Use stealable task placement when post-I/O CPU work is dominant and can benefit from migration.
• As a practical starting heuristic: if active stream count is at least 2x shard count and streams are long-lived, prefer round-robin/distributed mode.

Engineering Method

Development style is red/green TDD:

1. Add failing tests.
2. Implement minimal passing behavior.
3. Validate with full test and benchmark checks.

License

This project is licensed under the MIT License. See LICENSE.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in spargio by you shall be licensed as MIT, without any additional terms or conditions.