Presentation
Optimizing the Performance of NDP Operations by Retrieving File Semantics in Storage
DescriptionThe Near-Data Processing (NDP) architecture offloads computing to the storage device to avoid data movement.
However, before a NDP task is issued, the user library often needs to obtain the file semantics through relevant I/O stacks in the kernel, which incurs multiple intermediate data transfer overhead.
In this paper, we propose to ``File Semantics Retriever" in storage, which can directly provide file semantics to NDP modules and optimize data provisioning. The experimental results show that our solution is up to 83.94\% faster than Kernel interfaces in resolving file layouts and helps improve the performance of NDP operations.
However, before a NDP task is issued, the user library often needs to obtain the file semantics through relevant I/O stacks in the kernel, which incurs multiple intermediate data transfer overhead.
In this paper, we propose to ``File Semantics Retriever" in storage, which can directly provide file semantics to NDP modules and optimize data provisioning. The experimental results show that our solution is up to 83.94\% faster than Kernel interfaces in resolving file layouts and helps improve the performance of NDP operations.
Event Type
Research Manuscript
TimeTuesday, July 11th11:25am - 11:40am PDT
Location3008, 3rd Floor
Cloud
Embedded Systems
Embedded Memory, Storage and Networking