TL;DR
Linux I/O performance is about to get a massive boost after kernel developer Jens Axboe began hacking on new patches that deliver a 60% increase to per-core I/O performance, directly motivated by a presentation at the LSFMM summit in Croatia that exposed Linux's I/O overhead compared to the Storage Performance Development Kit (SPDK). This matters right now because it directly addresses a long-standing bottleneck in Linux storage performance, potentially reshaping how databases, cloud infrastructure, and high-performance storage systems operate.
What Happened
At last week's Linux storage, file-system, memory management and BPF summit (LSFMM) in Croatia, a presentation laid bare the stark I/O overhead gap between Linux and the Storage Performance Development Kit (SPDK) — and Jens Axboe, the maintainer of the Linux block layer, walked out of that room and started coding. Within days, he posted new patches that claim a 60% increase in per-core I/O performance, a leap that could fundamentally alter how Linux handles storage at scale.
Key Facts
- The presentation at LSFMM in Croatia compared Linux I/O overhead directly against SPDK, a user-space storage framework developed by Intel that bypasses the kernel entirely for maximum performance.
- Jens Axboe, the long-time maintainer of the Linux block I/O layer, was the developer who pursued these patches after the summit.
- The patches target per-core I/O performance, aiming for a 60% improvement in how many I/O operations each CPU core can process per second.
- The work is preliminary and has been posted for review on the Linux kernel mailing list, with no guarantee of immediate upstream acceptance.
- The performance gains come from reducing kernel-side overhead in the I/O path, particularly in how requests are submitted and completed.
- SPDK, which runs entirely in user space, has long outperformed Linux's kernel-based I/O stack for raw NVMe drives, often by 2–3x or more in per-core throughput.
- Axboe's patches are specifically designed to narrow that gap without abandoning the kernel's inherent advantages in security, isolation, and hardware abstraction.
Breaking It Down
The core tension here is between two competing philosophies of storage performance: kernel-based I/O, which prioritizes safety and abstraction, and user-space I/O, which prioritizes raw speed. SPDK has been the poster child for the latter, achieving absurdly high per-core I/O rates by stripping away every layer of kernel mediation. Linux's block layer, by contrast, has accumulated decades of compatibility layers, scheduling logic, and security checks. The LSFMM presentation quantified exactly how much that overhead costs.
The presentation at LSFMM showed that for a single NVMe drive, Linux's kernel I/O path can consume over 50% more CPU cycles per I/O operation compared to SPDK's user-space approach, with the gap widening as queue depths increase.
This is not a theoretical problem. In cloud environments where each CPU core is a billable resource, that overhead translates directly into higher costs and lower throughput for databases, object stores, and analytics workloads. Companies running MySQL, PostgreSQL, RocksDB, or SPDK-accelerated storage appliances have felt this pain for years. Axboe's patches target the most egregious sources of overhead: redundant locking in the submission path, unnecessary memory barriers, and inefficient completion handling. By streamlining these hot paths, he aims to reclaim a 60% performance uplift without requiring users to abandon the kernel's security model.
This is not the first time Axboe has tackled this problem. He previously introduced io_uring, a Linux kernel interface for asynchronous I/O that itself delivered major performance gains over the older AIO and sync I/O approaches. But io_uring, while transformative, still ran through the kernel — it just reduced syscall overhead. The new patches go deeper, attacking overhead within the block layer itself. The 60% figure is particularly striking because it applies to per-core performance, meaning that scaling out across many cores should yield multiplicative gains. For a 96-core server running a storage workload, a 60% per-core improvement could translate into a near-doubling of total system throughput under ideal conditions.
What Comes Next
- Patch review and iteration on the Linux kernel mailing list (May–June 2026): The patches are preliminary and will undergo rigorous review from other kernel maintainers. Expect multiple revisions as edge cases and regressions are identified.
- Potential inclusion in Linux 6.12 or 6.13 (late 2026): If the patches mature quickly, they could land in the next development cycle. Axboe's reputation and the clear performance motivation increase the odds of fast-tracking.
- Benchmarking from cloud providers and storage vendors (Q3 2026): Companies like Amazon (AWS Nitro), Google (GCP), and Pure Storage will likely run their own tests to validate the 60% claim on real hardware and production workloads.
- Competitive response from SPDK community (2026–2027): If Linux closes the gap significantly, the SPDK project may respond with further optimizations or new features to maintain its performance lead.
The Bigger Picture
This story sits at the intersection of two major trends: kernel-bypass vs. kernel-native performance and the commoditization of high-performance storage. For years, the industry has moved toward kernel-bypass techniques — DPDK for networking, SPDK for storage, and GPUDirect for accelerators — because the kernel was seen as a performance bottleneck. But maintaining separate user-space stacks is expensive and creates fragmentation. Axboe's work suggests that the kernel can be made competitive again, potentially reversing that trend.
At the same time, NVMe and persistent memory have made raw storage devices so fast that software overhead now dominates. A single modern NVMe drive can handle millions of IOPS, but the kernel's I/O stack often caps out at a few hundred thousand per core. The 60% improvement directly attacks this software bottleneck, which is the last major barrier to fully utilizing next-generation storage hardware. If successful, these patches could extend the life of kernel-based storage for another generation of hardware.
Key Takeaways
- [Performance Leap]: Jens Axboe's new Linux patches target a 60% increase in per-core I/O performance, directly motivated by a comparison against SPDK at the LSFMM summit in Croatia.
- [Kernel vs. User-Space]: The patches aim to close the gap with SPDK without abandoning the kernel's security and abstraction benefits, addressing a long-standing pain point for high-performance storage.
- [Timeline]: The patches are preliminary and under review; inclusion in a mainline Linux kernel release is likely months away, with potential arrival in late 2026.
- [Industry Impact]: Cloud providers, database operators, and storage vendors stand to benefit significantly if the 60% figure holds up in production, potentially reducing CPU costs for I/O-intensive workloads.
