고성능 컴퓨터 시스템 연구실

Device orchestration [MICRO’15, ISCA’18]. Modern high-performance servers aim to achieve high performance by exploiting a large number of high-performance devices (e.g., NVMe SSDs, 100GbE NICs, GPUs, AI accelerators) installed in a single server. Such examples include datacenter servers, SSD-array storage systems, and accelerator-array AI training servers. But, such servers suffer from the host-CPU’s high bandwidth requirement and device-control latency for orchestrating a large number of fast devices simultaneously. Our DCS architecture effectively resolves the problem by introducing DCS-Engine, a fast and scalable FPGA-based device orchestration device. On behalf of the host-side CPU and OS, DCS Engine manages all data and control paths among devices at the hardware level. In this way, a DCS server can significantly reduce the latency of inter-device communications, while bypassing the host CPUs mostly. In addition, by providing near-device processing (NDP) units on DCS Engine, a DCS server can further improve the overall performance and scalability. Storage server [HPCA’19, MICRO’19]. Modern storage servers aim to achieve high performance by deploying a number of SSDs on a single server. But, their performance, scalability, and cost-effectiveness highly depend on the cost of host CPUs and the data storage devices’ data reduction rate. To resolve the issues, we build CIDR and FIDR which are two DCS-inspired, fast, and scalable SSD-array supporting storage server architectures. Our DCS-inspired storage servers (1) offload the NDP-friendly operations from the host CPUs to FPGAs and (2) have them cooperatively manage the storage devices and deduplicate the data. AI training server [MICRO’20]. Modern AI training servers aim to achieve high training performance by deployin