Key Innovations from FAST ‘26 Full Papers

This report is generated by the OpenAI Deep Research.

Here, There and Everywhere: The Past, the Present and the Future of Local Storage in Cloud (from Shanghai Jiao Tong University, Alibaba Group, Solidigm)

Main Innovation

This paper presents a real-world evolution of cloud local storage architectures at Alibaba Cloud and proposes a new hybrid design for future systems. The authors analyze three generations of production systems—ESPRESSO, DOPPIO, and RISTRETTO—each addressing limitations of the previous design as SSD performance rapidly increased. ESPRESSO moves the storage stack from the kernel into user space using SPDK and polling to reduce context-switch overhead and improve NVMe utilization. DOPPIO then offloads virtualization and I/O handling to ASIC-based DPUs, eliminating host CPU overhead and enabling bare-metal compatibility. Finally, RISTRETTO introduces a hardware/software co-design combining ASIC acceleration with a programmable ARM-based SoC, allowing high performance while maintaining flexibility for cloud features.

Building on these lessons, the paper proposes LATTE, a hybrid storage architecture that integrates local NVMe storage with elastic block storage (EBS). LATTE uses machine-learning–based I/O dispatching and cache admission policies to dynamically decide whether requests should be served by fast local disks or remote block storage. This approach preserves the near-physical performance of local disks while improving availability and elasticity—two traditional weaknesses of local storage in cloud environments.

Academic Significance

The work provides a rare longitudinal study of production cloud storage architecture, offering insights into how storage stacks must evolve alongside hardware changes such as the rapid growth in NVMe SSD performance. It demonstrates how software-only approaches, hardware offloading, and hybrid co-designed architectures each address different bottlenecks (e.g., context switching, CPU utilization, and virtualization overhead). By analyzing the trade-offs between these designs in real deployments, the paper contributes empirical knowledge about the interaction between operating systems, virtualization layers, and modern storage devices.

Academically, the paper also highlights an emerging design direction: hardware–software co-design for storage virtualization. The RISTRETTO architecture shows how combining programmable SoC components with specialized ASIC logic can achieve both high performance and feature flexibility. The LATTE hybrid model further suggests new research opportunities in intelligent storage tiering, particularly using machine learning to manage heterogeneous storage resources in cloud environments.

Industry Significance

For cloud providers, the paper demonstrates how storage architecture must evolve to fully exploit modern NVMe SSD performance while keeping operational costs manageable. The transition from software stacks (ESPRESSO) to hardware-offloaded DPUs (DOPPIO) and then to hybrid ASIC/SoC platforms (RISTRETTO) shows practical strategies for scaling IOPS and throughput while reducing CPU overhead and improving resource efficiency. In production deployments, these designs can deliver millions of IOPS per node and near physical disk performance for virtual machines.

The proposed LATTE architecture has direct implications for real cloud platforms. By combining local disks with networked block storage, providers can offer both high performance and high availability without significantly increasing cost. This hybrid approach is particularly relevant for modern workloads such as large-scale analytics and LLM systems, which require high I/O performance but also elastic scaling and reliability. As a result, the design points toward future cloud storage services that dynamically blend local and disaggregated storage resources.