OpenCXD : An Open Real-Device-Guided Hybrid Evaluation Framework for CXL-SSDs
Hyunsun Chung 1,∗ , Junhyeok Park 1,∗ , Taewan Noh 1 , Seonghoon Ahn 1 Kihwan Kim 1 , Ming Zhao 2 , Youngjae Kim 1,† 1 Sogang University, Seoul, Republic of Korea, 2 Arizona State University, Tempe, AZ, USA · 2025
Read OpenCXD : An Open Real-Device-Guided Hybrid Evaluation Framework for CXL-SSDs by Hyunsun Chung 1,∗ , Junhyeok Park 1,∗ , Taewan Noh 1 , Seonghoon Ahn 1 Kihwan Kim 1 , Ming Zhao 2 , Youngjae Kim 1,† 1 Sogang University, Seoul, Republic of Korea, 2 Arizona State University, Tempe, AZ, USA (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-11477.txt.
Introduction
Hyunsun Chung and six co-authors present a test rig for a new kind of memory device. Their paper from twenty twenty-five concerns SSD drives that join the memory ranks through Compute Express Link, or CXL. These drives promise large space with memory-like access. Few real devices exist, so tests have leaned on software models.
Models miss how real firmware acts inside the drive. The team's rig, OpenCXD, joins a fine host-side model with a real open SSD board. The board runs true firmware while fake memory calls flow in. This mix shows effects that pure models hide. The authors say the rig can guide firmware design for these drives. It gives builders a closer look at speed and conduct before full hardware ships.
Catalog record
| CREATOR | Hyunsun Chung 1,∗ , Junhyeok Park 1,∗ , Taewan Noh 1 , Seonghoon Ahn 1 Kihwan Kim 1 , Ming Zhao 2 , Youngjae Kim 1,† 1 Sogang University, Seoul, Republic of Korea, 2 Arizona State University, Tempe, AZ, USA |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-11477
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
OpenCXD : An Open Real-Device-Guided Hybrid Evaluation Framework for CXL-SSDs Hyunsun Chung 1,∗ , Junhyeok Park 1,∗ , Taewan Noh 1 , Seonghoon Ahn 1 Kihwan Kim 1 , Ming Zhao 2 , Youngjae Kim 1,† 1 Sogang University, Seoul, Republic of Korea, 2 Arizona State University, Tempe, AZ, USA Abstract The advent of Compute Express Link (CXL) enables SSDs to participate in the memory hierarchy as large-capacity, byte-addressable memory devices. These CXL-enabled SSDs (CXL-SSDs) offer a promising new tier between DRAM and traditional storage, combining NAND flash density with memory-like access semantics. However, evaluating the performance of CXL-SSDs remains difficult due to the lack of hardware that natively supports the CXL.mem protocol on SSDs. As a result, most prior work relies on hybrid simulators combining CPU models augmented with CXL.mem semantics and SSD simulators that approximate internal flash behaviors. While effective for early-stage exploration, this approach cannot faithfully model firmware-level interactions and low-level storage dynamics critical to CXL-SSD performance. In this paper, we present OpenCXD , a real-device-guided hybrid evaluation framework that bridges the gap between simulation and hardware. OpenCXD integrates a cycle-accurate CXL.mem simulator on the host side with a physical OpenSSD platform running real firmware. This enables in-situ firmware execution triggered by simulated memory requests. Through these contributions, OpenCXD reflects device-level phenomena unobservable in simulation-only setups, providing critical insights for future firmware design tailored to CXL-SSDs. I Introduction The growing scale of modern deep learning models and data analytics applications has led to memory footprints reaching tens of terabytes [ 1 , 2 , 3 ] , far beyond the limits of traditional DRAM installations. This widening gap between demand and capacity has resurrected the infamous memory wall [ 4 ] , in which memory bandwidth and size become critical bottlenecks to system performance. To mitigate this issue, researchers have begun exploring memory expansion techniques that repurpose alternative technologies as additional memory [ 5 , 6 , 7 ] , paving the way for new architectural paradigms. Among these, Compute Express Link (CXL) [ 8 ] has emerged as a promising enabler of such memory expansion. CXL is a high-bandwidth, cache-coherent interconnect built on top of the PCI Express (PCIe) [ 9 ] infrastructure. By using CXL, large-capacity PCIe devices like flash-based Solid-State Drives (SSDs) can be attached directly to the host system memory space, creating a memory pool that augments or disaggregates DRAM [ 10 ] . Notably, CXL’s memory semantics (CXL.mem) support byte-addressable access to device memory, meaning a CPU can read or write an SSD’s onboard DRAM buffer with ordinary load and store instructions. This eliminates the need for traditional I/O commands, significantly reducing software overhead and access latency while enabling a unified, tiered memory architecture that extends beyond DRAM’s capacity limits [ 11 ] . Building on this capability, a new class of memory-semantic devices, CXL-enabled SSDs (CXL-SSDs), has begun to emerge (§ II ). These devices leverage mature NAND flash technology to offer terabytes of byte-addressable capacity at a fraction of DRAM’s cost per gigabyte [ 7 ] , while maintaining access latencies in the microsecond range. Although slower than DRAM by several orders of magnitude, CXL-SSDs provide significantly lower latency than traditional SSDs accessed via the block interface (e.g., NVMe [ 12 ] ). The key challenge for CXL-SSDs is how to architect and evaluate these devices effectively, maximizing their performance potential while addressing inherent latency trade-offs. However, this evaluation remains difficult in practice, due to the lack of hardware that natively supports the CXL.mem protocol on SSDs . To overcome this, recent research has adopted software-based hybrid simulators, combining CPU simulators (e.g., MacSim [ 13 ] , Gem5 [ 14 ] ) extended with CXL.mem semantics and SSD simulators (e.g., SimpleSSD [ 15 ] , FlashSim [ 16 ] ) that model internal flash behaviors such as address translation and I/O scheduling. This methodology enables early-stage design exploration and has been widely used in prior work [ 17 , 18 , 19 ] . Notably, the CXL.mem interface overhead itself has been characterized in prior studies [ 20 , 21 ] , and shown to be relatively consistent and bounded [ 10 ] . As such, injecting this CXL interface time overhead into x86 simulations as a parameter is generally considered a reasonable approach for modeling host-side access costs. However, the same cannot be said for modeling device-side behavior . Replacing a real SSD with a simulator introduces significant challenges in capturing the full complexity of CXL-SSD-specific firmware logic and storage interactions, which are limitations that fundamentally hinder accurate evaluation of CXL-SSD designs (§ III ). There are two key limitations of simulation-only evaluation. First , since CXL-SSDs function as memory rather than storage, they must handle fine-grained, cacheline-level memory accesses, making them highly sensitive to device-side performance fluctuations. However, simulators typically rely on static latency models, overlooking dynamic behaviors such as real-time NAND latency variability and firmware delays, resulting in latency estimation errors as high as 36% [ 22 ] . Second , CXL-SSDs introduce new firmware-managed mechanisms, such as write logging and log compaction [ 11 ] , that do not exist in conventional SSDs. Simulating these new behaviors solely from the host side, without executing them on real hardware and capturing internal device interactions, fails to reflect crucial contention and scheduling effects, yielding an incomplete and often misleading performance picture [ 23 ] . In this paper, we present OpenCXD , a real-device-guided hybrid evaluation framework for CXL-SSDs. OpenCXD combines a cycle-accurate CPU and memory simulator augmented with CXL.mem support and a hardware platform based on OpenSSD [ 23 ] , an open-source SSD prototype that runs actual controller firmware. In essence, the host side of the system (CPU, cache, and memory controllers) is simulated, allowing full control over experimental scenarios and observability of system-level events, while the storage side is handled by a physical device that embodies a CXL-SSD (§ IV ). We addresses two key challenges in developing OpenCXD : \bullet Enabling Cacheline-Level Access over NVMe. Unlike DRAM or DRAM-based CXL memory modules [ 24 ] , publicly available SSD prototypes do not natively support cacheline-sized CXL.mem transactions. Bridging this gap is non-trivial because NVMe operates on fixed-sized (e.g., 4 KB) blocks with DMA-based transfers, not on byte-addressable memory [ 25 , 26 ] . To address this, OpenCXD defines custom NVMe commands that encode CXL.mem semantics and emulate cacheline-granularity memory access without modifying the physical hardware interface. \bullet Cycle-Level Timing Integration with Real Firmware. While the cycle-accurate host simulator tracks timing at the cycle level, the OpenSSD operates asynchronously with real-time firmware execution. OpenCXD resolves this by adopting a device-in-the-loop design: for each memory access, the simulator pauses execution, delegates the operation to OpenSSD running real firmware, and waits until the firmware measures the end-to-end latency and reports it to the host simulator. The simulator then resumes by converting the measured latency to cycles using a calibrated timing ratio and advances its internal clock accordingly. We then implement key firmware components of a state-of-the-art CXL-SSD [ 11 ] within the OpenSSD device. This allows fine-grained memory requests from the host to trigger CXL-SSD firmware execution. These design choices enable OpenCXD to maintain the configurability