Understanding UFS vs NVMe Software Stack Overhead in Linux Systems

Storage performance has become a critical component in modern computing, especially in embedded systems, smartphones, and laptops. Two major storage technologies—Universal Flash Storage (UFS) and Non-Volatile Memory Express (NVMe)—are at the center of this evolution. While both offer high performance and low latency, their software stack overheads vary significantly, especially in Linux-based systems.

This blog dives deep into the UFS and NVMe storage software stacks on Linux, explores system-level overheads, and highlights findings from a detailed performance comparison presentation by Micron.

1. Overview of UFS and NVMe Software Stacks in Linux

UFS (Universal Flash Storage)

UFS is a JEDEC standard primarily used in smartphones and embedded systems. It offers:

• High-speed serial interface (MIPI M-PHY + UniPro)
• Full-duplex communication
• Command queuing (native NCQ support)
• Low power consumption

UFS in the Linux Kernel

• First-class support since Linux kernel 3.18
• Modern devices use kernel 5.x and above with enhanced UFS features

• Key Linux components for UFS:

  • SCSI UFS driver (ufshcd.c)
  • SCSI mid-layer
  • Block layer (blk-mq)
  • I/O scheduler (None, MQ-Deadline, BFQ, etc.)

UFS Stack Flow in Linux

User I/O Request
      ↓
Filesystem (ext4, f2fs, etc.)
      ↓
Block Layer (blk-mq)
      ↓
SCSI Mid-layer
      ↓
UFS Host Controller Driver (`ufshcd`)
      ↓
UniPro/M-PHY Interface
      ↓
UFS Device

NVMe (Non-Volatile Memory Express)

NVMe is designed specifically for SSDs connected over PCIe. It is widely used in:

• Laptops and desktops
• Data centers
• Removable SSDs (e.g., USB4 NVMe enclosures)

NVMe in the Linux Kernel

• Introduced in Linux 3.3, matured in 4.x+ series

• Key Linux components for NVMe:

  • NVMe driver (nvme-core, nvme.c)
  • Block layer (blk-mq)
  • Direct I/O submission to hardware

NVMe Stack Flow in Linux

User I/O Request
      ↓
Filesystem
      ↓
Block Layer (blk-mq)
      ↓
NVMe Driver (`nvme.c`)
      ↓
PCIe
      ↓
NVMe Device

Key NVMe Advantages

• No SCSI overhead (unlike UFS)
• Simpler protocol stack
• Lower software latency

2. Micron Presentation: UFS vs NVMe Storage Stack Overhead Comparison

In an insightful presentation, titled “UFS/NVMe Storage Stack System Level Performance in Embedded Systems”, the software overhead of UFS and NVMe stacks in embedded Linux systems was evaluated.

📄 Download the PDF Slides

Key Points from the Presentation:

• Test Environment:

  • Embedded system with both UFS and NVMe storage
  • Linux Kernel 5.4/5.10 (depending on platform)
  • Identical workloads across both devices

• Metrics Analyzed:

  • IOPS (Input/Output Operations Per Second)
  • Latency per layer (block layer, driver layer, etc.)
  • CPU usage and context switch rate
  • Interrupt count and scheduling cost

• Findings:

  • NVMe has 30-50% lower software stack latency than UFS in random read workloads.
  • UFS software stack introduces extra overhead due to SCSI mid-layer and controller complexity.
  • The Linux SCSI stack is a legacy burden for UFS, not present in NVMe path.
  • In high-concurrency scenarios, NVMe handles scaling better due to its native multiqueue support and lightweight protocol.

• Power Efficiency:

  • UFS is still more power-efficient for mobile usage, justifying its dominance in smartphones.
  • NVMe consumes more power but provides higher throughput per watt in high-performance systems.

Supporting Resources:

• Blog: Linux Storage System Bottleneck for eMMC/UFS
• Legacy Slides: Linux Storage System Bottleneck Exploration (PDF)

3. Industry Usage Scenarios

4. Conclusion

The choice between UFS and NVMe in Linux systems isn’t just about hardware performance—it heavily depends on software stack efficiency. NVMe offers a cleaner, faster path, while UFS lags due to legacy SCSI components. However, UFS remains relevant in power-sensitive applications like mobile devices.

As Linux continues to evolve (especially with features like IO_uring, blk-mq optimizations, and lightweight SCSI layering), we may see these gaps narrow. For now, NVMe is the clear winner in low-overhead, high-performance environments, while UFS remains the choice for mobile-centric designs.

Author Note: This blog is based on public presentations and research by Bean Huo and is intended for educational and reference use for developers and system integrators.