Jianshen Liu: Eusocial Storage: Embedded Storage Devices that Act Collectively

June 18, 2020

By , Assistant Director 

Eusocial insects working together
Jianshen Liu

Event Information:

Date and time: June 23, 2020, 10-11am
Zoom link

With the end of Dennard Scaling and Moore’s Law, speed increases of CPU and DRAM are falling behind speed increases of storage and network devices, making traditional storage systems architectures inadequate. First, when buying storage devices, users traditionally have to buy host infrastructure to use these devices. This "tax" of data access increases as the performance growth of storage devices outpaces that of the traditional computing resources (e.g., CPU and DRAM). Second, the bandwidth of storage device interconnects and network fabrics for remote data access are becoming the bottleneck of accessing data from high-performance storage devices.

To solve these problems, storage devices need to be more autonomous (less dependent on host infrastructure) and support bandwidth-reducing functions to be executed where the data resides, i.e., offloading these functions to storage devices themselves. We propose a storage architecture that allows embedded storage devices to autonomously carry out different storage services offloaded from hosts by enabling collective action of these devices. We call this approach "eusocial storage devices" as it was inspired by the collective behaviors of eusocial insects like ants and termites. This architecture enables several classes of function offloading to storage devices for services that are traditionally running on server-based storage systems. Offloading functions to storage devices could reduce the tax and lower the bandwidth requirement of data access. Furthermore, this architecture presents strong flexibility to leverage existing and upcoming heterogeneous/diverse hardware resources to boost data access performance for different applications.

In the proposed project, we will explore the design space of eusocial storage devices, including quantifying the cost-benefit of offloading functions to storage devices, evaluating different architectures, and identifying opportunities for leveraging hardware accelerators/specialized hardware.

The proposed project builds on our earlier work introducing the "media-based work unit" (MBWU, pronounced "MibeeWu") to quantify the cost-benefit of offloading a given data access function to a specific storage device, and on our quantification of the benefit of using eusocial storage devices over traditional storage devices at edge data centers, where on-site maintenance cost is of critical importance.