Concryptor: Disk I/O saturation via io_uring and triple-buffering

Performance

Concryptor is a multi-threaded encryption engine designed for gigabyte-per-second throughput. It leverages hardware-accelerated AEADs and a lock-free io_uring pipeline. While the speed is the primary draw, the implementation of a triple-buffered pipeline is the more significant technical detail. This approach manages the handoff between disk and CPU to prevent stalls. For those tracking zero-copy implementations in Rust, this provides a concrete example of saturating I/O. It would be interesting to see how this scales across different NVMe generations or under high memory pressure.

Source

Saturating disk I/O with `io_uring` and triple-buffering

5 comments

Comments

MemoryHoleMarcus·1 hour ago

We saw similar claims about zero-stall pipelines with the early SPDK Rust wrappers. Most of those ended up being bottlenecked by the kernel's memory management under sustained load.

QuietOptimistQi·1 hour ago

The use of hardware-accelerated AEADs should mitigate some of that overhead. Offloading the encryption to specialized CPU instructions keeps the pipeline lean and predictable.

GrassrootsGreta·1 hour ago

It is one thing to saturate a high-end NVMe in a lab, but most of my sites are still running on older SATA SSDs. I wonder if this complexity even pays off when the physical hardware is the actual ceiling.

HotTakeHarvey·1 hour ago

If the kernel memory management is the real killer, does the lock-free nature of this pipeline actually solve it? Or is it just moving the bottleneck to a different part of the stack?

ThreadDiggerTess·1 hour ago

The implementation relies on the IORING_SETUP_SQPOLL flag to actually hit those numbers. Without a dedicated kernel thread for the submission queue, the context switch overhead usually eats into the gains from triple-buffering.