SkyByte: Architecting an Efficient Memory-Semantic CXL-based SSD with OS and Hardware Co-design

Haoyang Zhang, Yuqi Xue, Yirui Eric Zhou, Shaobo Li, Jian Huang

Introduction

Compute Express Link (CXL)-based SSDs offer a highly cost-effective way to expand server main memory, but their long flash latencies and page-level access granularities severely bottleneck CPU performance. SkyByte solves this through a clever OS-hardware co-design that introduces coordinated context switching to hide latency, alongside a cacheline-granular write log to minimize flash I/O. The result is a 6x performance boost and a massive 23x reduction in write traffic, making SSDs a truly viable memory expansion tier.

The Challenge: Memory-Semantic SSD Bottlenecks

The CXL standard allows CPUs to access connected devices—like SSDs—directly via standard memory load/store instructions. This creates "memory-semantic SSDs" that act as cheap, massive extensions of main memory. However, treating an SSD exactly like DRAM out-of-the-box leads to disastrous performance due to two main challenges:

  1. Massive Latency and CPU Stalls: Flash memory is orders of magnitude slower than DRAM. When a CPU requests data that isn't in the SSD's internal DRAM cache, the processor pipeline stalls for microseconds. Because these are direct memory instructions, the OS is unaware of the stall and cannot context-switch to another thread, leaving the CPU core completely idle.
  2. Granularity Mismatch: CXL operates at cacheline granularity (64 bytes), but flash chips read and write in large pages (e.g., 4KB). Caching entire 4KB pages in the SSD's limited internal DRAM when the CPU only needs 64 bytes wastes precious cache space and causes severe write amplification, wearing out the SSD prematurely.
Normalized execution time comparison between DRAM memory and baseline CXL-SSD across various workloads.

SkyByte: A Holistic Co-Design Approach

SkyByte tackles these bottlenecks using a holistic co-design between the host OS and the SSD controller to bridge the gap between byte-addressable CPU requests and block-oriented flash storage.

SkyByte system memory mapping (a) and internal hardware architecture (b) featuring a DRAM cache and cacheline-granular write log.

Coordinated Context Switch

SkyByte bridges the communication gap between the SSD and the host CPU. When the SSD controller detects that a memory request will take a long time (e.g., a cache miss requiring a flash read), it sends a custom CXL "No Data Response" message with a special SkyByte-Delay opcode. The host CPU translates this into a hardware exception, allowing the OS scheduler to instantly context-switch to another runnable thread instead of stalling.

CXL-Aware SSD DRAM Management

To fix the granularity mismatch, SkyByte re-architects the SSD's internal DRAM. Instead of just caching 4KB pages, it splits the memory into a cacheline-granular write log and a page-granular read-write cache. Small writes are appended directly to the log, absorbing the I/O and coalescing multiple small writes into single page writes in the background. This prevents cache thrashing and reduces write amplification.

Adaptive Page Migration

SkyByte tracks the "hottest" (most frequently accessed) pages in the SSD and transparently promotes them to the faster host DRAM using PCIe interrupts, ensuring the most critical data runs at native memory speeds.

Key Results

The authors evaluated SkyByte using multi-threaded instruction traces from standard data-intensive benchmark suites replayed in a cycle-accurate simulator. Key findings include:

  • Massive Performance Gains: SkyByte outperforms state-of-the-art CXL-SSD baselines by an average of 6.11x (and up to 16.35x on certain workloads).
  • Near-Ideal Speed at Lower Cost: SkyByte achieves 75% of the performance of an ideal "infinite host DRAM" setup. Given the price difference between DDR5 and flash memory, the authors estimate SkyByte improves cost-effectiveness by 11.8x.
  • Drastically Reduced Wear: The cacheline-granular write log reduces flash write traffic by an average of 23.08x, which significantly extends the lifespan of the SSD.
  • Lower Latency: Average memory access time (AMAT) is reduced by 14.19x compared to baseline CXL-SSDs.

Conclusion

SkyByte demonstrates that CPUs cannot hide flash latency alone; involving the OS scheduler via hardware exceptions is mandatory for efficient CXL-SSD integration. By combining latency hiding (context switching) with latency reduction (write logging and page migration), SkyByte transforms SSDs into a high-performance memory expansion tier. Future work will focus on native data persistence guarantees and optimizing for complex multi-socket NUMA architectures.

Summary updated

BibTeX
@inproceedings{zhang2025skybyte,
  author = {Haoyang Zhang and Yuqi Xue and Yirui Eric Zhou and Shaobo Li and Jian Huang},
  title = {{SkyByte: Architecting an Efficient Memory-Semantic CXL-based SSD with OS and Hardware Co-design}},
  booktitle = {HPCA},
  year = {2025},
  url = {https://openreview.net/forum?id=6nilRTBMwg},
}
Figure