Authors: Soon Hwang (Sogang University); Junhyeok Park (Sogang University, SK Hynix); Junghyun Ryu and Seonghoon Ahn (Sogang University); Jeoungahn Park, Jeongjin Lee, Jinna Yang, Soonyeal Yang, and Jungki Noh (SK Hynix); Qing Zheng (Los Alamos National Laboratory (LANL)); Woosuk Chung and Hoshik Kim (SK Hynix); and Youngjae Kim (Sogang University)
Abstract: Scientific analytics on disaggregated object storage suffer from high data movement overhead, often dominating query latency. While computation-enabled object storage (COS) reduces this cost via operator offloading, existing approaches assume a fixed execution boundary and restrict execution to a single storage-side layer, leaving cross-layer data movement unoptimized. We observe that this stems from a mismatch between operator semantics and execution scope, where early reduction and globally coordinated execution require different layers. Therefore, the key challenge is not operator selection, but execution boundary placement across storage layers.
We present OASIS, the first COS system that formulates storage-side query execution as a cross-layer boundary placement problem over a multi-layer hierarchy. Its core algorithm, SODA, determines the split point to minimize intermediate data movement. Integrated with Apache Spark, OASIS reduces internal data movement by 99.998% and lowers end-to-end query latency by up to 25% compared to Skyhook, a state-of-the-art COS system.
Presentation: file