Esmond Chang

Developed a solution for an AI recommendation system to manage Terabyte-sized datasets that are difficult to fit into single server memory, causing higher costs and search latency. Optimized the ANN (approximate nearest neighbor) library to leverage Persistent Memory's capacity and performance, fitting the Terabyte dataset into a single server while delivering the same recall performance with 30% lower TCO compared to DRAM solutions.

Designed and deployed EBS (Elastic Block Storage) for CSPs using hundreds of Petabytes of QLC drives on Intel Xeon servers. Utilized SPDK (Storage Performance Development Kit) FTL and Persistent Memory as a buffer to resolve the performance and endurance bottlenecks of QLC drives. The solution leverages Persistent Memory with microsecond write latency and high endurance over 100-DWPD to reduce overall OPEX by 40%.

Addressed traditional NAS low performance in synchronous write with slow media backend storage (HDD) and enhanced read latency with a larger capacity of the read cache. Leveraged Persistent Memory as ZFS ZIL Slog buffer and L2ARC cache. This approach improved random write performance by 1.3X and reduced read latency by 98%.

Developed a solution for modern applications on microservices with stateful database storage, which could be a performance bottleneck in scalability and availability. The solution leverages Persistent Memory's low latency and high capacity as container attached storage with CSI (Container Storage Interface) for workloads such as relational databases and AI training, improving OLTP transaction performance (tps) for mission-critical microservice-based applications.

Overcame traditional hardware RAID limitations on PCIe bandwidth to scale NVMe SSDs and increased configuration complexity for scalability. A software RAID based on the mdadm open-source solution provides flexibility in RAID configuration without PCIe lane x16 bandwidth limitations. Intel VROC (Virtual RAID-on-CPU) solution offers a 50% performance gain in RAID5 Random mix workload with lower CapEx as a replacement for Hardware HBA.

Managed surprise hot-plugging complexities when the NVMe SSD controller is located inside the SSD. Intel VMD (Volume Management Device), a feature inside Xeon CPUs, works with software as NVMe SSD endpoint root port to enhance manageability and surprise hot-plug stability. The VMD solution provides rich features such as error isolation, LED management, and bootable RAID, supporting Linux, Windows, and HCI environments like VMware in Xeon CPUs for enterprise and CSP customers.