Glommio

Glommio is the most architecturally ambitious thread-per-core Rust runtime, designed by Glauber Costa at DataDog after seven years building ScyllaDB's seastar framework. It features a triple io-uring ring design, proportional-share scheduling with latency classes, and direct-io as a first-class citizen — capabilities no other Rust runtime offers. It is also effectively unmaintained since Costa's departure to co-found Turso in 2021.

Triple-ring architecture

Each thread operates three independent io_uring instances:

• Main Ring — throughput-oriented operations.
• Latency Ring — time-sensitive tasks; a timer ensures yield within a specified duration.
• Poll Ring — NVMe operations using io_uring's interrupt-free polling mode.

Before the main ring blocks, the latency ring's file descriptor is registered for poll on the main ring, ensuring latency-sensitive completions wake sleeping threads. Checking the latency ring costs only two integer comparisons (CQ head and tail pointers), making yield_if_needed() nearly free.

Proportional-share scheduling

Task queues have configurable Shares: if queue A has 2 shares and queue B has 1, A receives 2/3 of CPU time under contention. Shares can be static or dynamic, adjusted at runtime by embedded controllers that observe metrics and auto-tune weights — a feedback loop for self-tuning storage systems. Each queue declares a Latency class: NotImportant (throughput-optimized) or Matters(Duration) (latency-sensitive, with io_uring timer-driven preemption). Default preemption timeslice is 100ms. A stall detection handler sends SIGUSR1 and collects stack traces when tasks exceed their expected runtime — observability neither monoio nor compio provides.

Storage I/O

Glommio's storage subsystem is unmatched. DmaFile provides random-access direct-io that bypasses the OS page cache entirely, allocating from io_uring's pre-registered buffer pool and returning reference-counted pointers for zero-copy device-to-user transfer. DmaStreamReader implements intelligent read-ahead that dispatches new reads as each buffer is consumed, maintaining constant I/O parallelism.

Costa's benchmarks on Intel Optane: sequential scans of a 53GB file (2.5x RAM) achieved 7.29 GB/s with Direct I/O vs 0.95 GB/s buffered — 7.7x improvement. Random reads hit 547K IOPS vs 238K IOPS buffered, using 80KB of memory versus gigabytes of page cache.

Buffer model trade-off

Unlike monoio's ownership-transfer model, Glommio copies data into internal managed buffers to provide a more traditional borrowing interface. ByteDance's benchmarks attribute Monoio's higher peak network throughput directly to this extra copy overhead.

Maintenance status

Effectively unmaintained since Costa's departure. Open issues accumulate slowly, and a memory corruption bug in the SPSC queue was reported December 2025. Teams considering Glommio should be prepared to fork and maintain it. The GMF framework demonstrated 84% more requests/sec than Tonic-based gRPC, proving the architecture works — the question is long-term viability.

When to use it

Only if you are building a storage engine that specifically needs proportional-share scheduling, DMA file abstractions, and latency-class differentiation — and you have the engineering capacity to maintain a fork. For everything else, compio is the recommended alternative.