Near-data processing (NDP) mitigates the data movement bottleneck in modern computing systems by performing computation closer to where the data resides. Solid-state drives (SSDs) are well-suited for NDP for four key reasons: (1) they store large application datasets that exceed main memory capacity; (2) they inherently support three near-data processing paradigms: processing using controller cores, processing using DRAM in the SSD, and in-flash processing; (3) they offer high computation throughput via massive internal parallelism; (4) they ease the burden on the memory hierarchy by supporting in-place computations.

Leveraging NDP techniques in SSDs presents three key challenges: (1) limited applicability across diverse workloads, (2) underutilization of SSD computation resources, and (3) significant programmer effort to analyze the applications and system state to select and integrate a suitable NDP technique in the storage I/O stack. While no prior approach leverages multiple SSD computation resources in a general-purpose and application-agnostic manner, several studies partition computation between the host and near-memory processing units. However, adapting these techniques to SSDs offers limited benefits because they (1) ignore the heterogeneity of the SSD computation resources, and (2) rely on limited factors such as bandwidth utilization or data movement cost, ignoring key factors such as computation resource utilization.

We address these challenges with Conduit, a general-purpose programmer-transparent NDP framework for SSDs that dynamically offloads computations across multiple SSD resources. Conduit performs: (1) a compile-time preprocessing step that vectorizes application code segments for NDP, and (2) a runtime instruction-granularity offloading step that dynamically determines the best-fit SSD computation resource based on resource utilization, data dependencies and data movement costs. Conduit accelerates a broad range of workloads without programmer intervention. We evaluate Conduit and two prior NDP offloading techniques using an in-house event-driven SSD simulator on six data-intensive applications. Conduit outperforms the best-performing prior offloading policy by 80% and reduces energy consumption by 46%, with minimal storage overhead, and no additional hardware costs.