NAND flash-based Solid-State Drives (SSDs) have become the predominant storage medium for modern data-intensive applications.
Compared to traditional Hard Disk Drives (HDDs), SSDs offer various advantages, including shorter I/O latency, compact form f...
NAND flash-based Solid-State Drives (SSDs) have become the predominant storage medium for modern data-intensive applications.
Compared to traditional Hard Disk Drives (HDDs), SSDs offer various advantages, including shorter I/O latency, compact form factor, and lower energy consumption.
Moreover, advances in multi-level cell technology and 3D wordline stacking have significantly increased SSD capacity, enabling a single device to store hundreds of terabytes of data.
With these benefits and remarkable capacity enhancements, SSDs have emerged as the de facto standard for modern data-intensive workloads.
While SSDs provide significantly higher performance than HDDs, their limited bandwidth still makes it challenging to meet the demanding performance requirements of data-intensive applications.
Modern data-intensive workloads demand extremely high data processing capability over massive datasets.
To satisfy these requirements, modern computing systems employ high-performance compute units (e.g., GPUs) capable of delivering hundreds of tera–floating-point operations per second.
Unfortunately, the bandwidth of SSDs, which physically store the data to be processed, remains limited to about 4–15 GB/s.
This severely constrained SSD bandwidth introduces a substantial performance bottleneck that cannot be addressed merely by increasing compute performance.
To overcome the limited bandwidth of SSDs, in-flash processing (IFP), which performs computation directly within flash chips inside the SSD, has been actively investigated.
Because the aggregated internal bandwidth across multiple flash chips within an SSD is higher than the external PCIe interface bandwidth, IFP can effectively increase the available bandwidth of SSDs.
Leveraging this advantage, several studies have proposed IFP-based acceleration techniques for various data-intensive applications.
IFP has the potential to deliver substantial performance benefits,
but several key challenges must be addressed before it can be adopted in practical systems.
First, achieving high bandwidth requires IFP to operate all flash chips concurrently, which exposes the SSD to significant read disturbance that severely degrades both its lifetime and performance.
Second, the capacity-oriented design trend of modern NAND flash memory has substantially reduced per-chip read bandwidth, directly undermining the performance benefits that IFP aims to deliver.
Third, flash chips are manufactured under extremely tight fabrication resource constraints, making the integration of additional IFP logic components both challenging and costly.
In particular, ensuring reliable IFP operation necessitates an on-chip ECC decoder, which incurs substantial area and power overhead.
In this dissertation, we propose two novel schemes that effectively address the key challenges hindering the practical adoption of IFP.
First, we introduce DEAR, a dynamic read-refresh scheme that mitigates the impact of the read disturbance by performing refreshes in a fine-grained manner.
Conventional SSD firmware ensures reliability against read disturbance by simply triggering a refresh once the number of reads to a block exceeds a predefined threshold.
However, due to the substantial process variation in modern 3D NAND flash memory, such coarse-grained refresh policies cause excessive and unnecessary data migrations.
To eliminate this inefficiency, DEAR performs read-refresh operations based on actual error rates rather than read counts, and operates at the wordline granularity instead of the block level.
By selectively refreshing only the highly disturbed wordlines, DEAR drastically reduces the data migrations during refreshes, thereby significantly improving both SSD lifetime and performance under read disturbance.
Second, we present AiF, an in-flash processing framework that effectively improves bandwidth and implementation overheads of IFP.
To overcome the limited performance benefits of IFP caused by the diminishing read bandwidth of flash chips, AiF employs a charge-recycling read technique that dramatically increases read bandwidth by reusing the voltages applied in previous read operations for subsequent reads.
In addition, to address the area and power overhead of on-chip ECC decoders required to correct the high error rate of NAND flash memory, AiF adopts biased error encoding, which intentionally introduces non-uniform error characteristics across pages to create highly reliable pages.
By selectively storing IFP data (e.g., LLM weights) only in these reliable pages, AiF ensures the reliability of IFP using lightweight ECC.
Combining these two techniques, AiF achieves up to 4x higher read bandwidth than conventional IFP while significantly reducing implementation overhead.
To evaluate the effectiveness of the proposed schemes, we conduct extensive experiments using real-world workloads and LLMs on SSD emulation platforms.
The evaluation results show that DEAR significantly reduces data copying caused by read-refresh operations, thereby substantially improving both SSD lifetime and performance for read-intensive workloads.
Furthermore, the IFP-specialized read and reliability techniques of AiF enable a dramatic increase in read bandwidth with marginal implementation overhead, leading to a significant improvement in LLM inference throughput.