Yussuf Khalil, Daniel Habicht, Pascal Ellinger, Frank Bellosa, Javier González, Adam Manzanares, Vivek Shah

DIMES'25, 3rd Workshop on Disruptive Memory Systems, Seoul, Korea

Abstract:

Upcoming CXL-based hybrid SSDs offer a persistent memory interface in addition to traditional block-based communication. Previous research argued that the non-uniform access latencies (on-device DRAM cache hits vs. misses) of these devices call for new OS-based resource management strategies as existing DAX mechanisms were not built with this device model in mind. In this work, we sketch Transparent DAX Mappings, a novel approach to enable a broad range of applications to benefit from hybrid SSDs. By opportunistically handing out DAX mappings to processes through a collaboration between the kernel and the C standard library, we aim to achieve automatic kernel bypass for unmodified I/O-heavy applications. We evaluate a preliminary implementation of our design on a Samsung CMM-H prototype and demonstrate up to 96.2 % increased throughput in Valkey.

BibTeX:

@inproceedings{10.1145/3764862.3768178,

author = {Khalil, Yussuf and Habicht, Daniel and Ellinger, Pascal and Bellosa, Frank and Gonz\'{a}lez, Javier and Manzanares, Adam and Shah, Vivek},

title = {Transparent DAX Mappings: Towards Automatic Kernel Bypass with CXL-Based Hybrid SSDs},

year = {2025},

isbn = {9798400722264},

publisher = {Association for Computing Machinery},

address = {New York, NY, USA},

url = {https://doi.org/10.1145/3764862.3768178},

doi = {10.1145/3764862.3768178},

abstract = {Upcoming CXL-based hybrid SSDs offer a persistent memory interface in addition to traditional block-based communication. Previous research argued that the non-uniform access latencies (on-device DRAM cache hits vs. misses) of these devices call for new OS-based resource management strategies as existing DAX mechanisms were not built with this device model in mind. In this work, we sketch Transparent DAX Mappings, a novel approach to enable a broad range of applications to benefit from hybrid SSDs. By opportunistically handing out DAX mappings to processes through a collaboration between the kernel and the C standard library, we aim to achieve automatic kernel bypass for unmodified I/O-heavy applications. We evaluate a preliminary implementation of our design on a Samsung CMM-H prototype and demonstrate up to 96.2 \% increased throughput in Valkey.},

booktitle = {Proceedings of the 3rd Workshop on Disruptive Memory Systems},

pages = {54–62},

numpages = {9},

keywords = {CXL, DAX, Hybrid SSDs, Kernel Bypass},

location = {Seoul, Republic of Korea},

series = {DIMES '25}

}