UnICom: A Universally High-Performant I/O Completion Mechanism for Modern Computer Systems

Modern computer systems are increasingly equipped with dozens to hundreds of cores, while high-performance Solid-State Drives (SSDs), enabled by NVMe and emerging technologies such as CXL-SSDs, provide massive I/O bandwidth and microsecond-scale latency. Yet, software overhead in the I/O stack remains a critical bottleneck, often contributing up to 50% of total I/O latency. Existing I/O completion mechanisms fall short: polling achieves low latency but wastes CPU cycles, whereas interrupts conserve CPU resources but incur significant wake-up overhead. This paper presents UnICom (Universal I/O Completion), a new I/O completion mechanism that unifies the benefits of polling and interrupts while avoiding their drawbacks. The key insight is that a kernel trap is negligible compared to disk I/O latency, yet enables access to kernel infrastructure for efficiency and security. Building on this, UnICom introduces three core techniques: TagSched, a lightweight tag-guided scheduling mechanism that minimizes sleep and wake-up overhead; TagPoll, a centralized kernel-level I/O completion thread that consolidates polling across threads and processes; and SKIP, a kernel-assisted direct-access mechanism that eliminates complex user-space permission management. Together, these techniques enable efficient multi-process support and direct SSD access while bypassing much of the kernel I/O stack. We implement UnI-Com in the Linux kernel and evaluate it against ext4, BypassD, and io_uring. Across all experiments, UnICom consistently delivers high I/O performance, matching or exceeding the best of polling and interrupts under both low and high CPU utilization.