The introduction of the NVIDIA BlueField-3 data processing units (DPUs) in mid- 2020 opened multiple possibilities to enhance systems architecture design. These DPUs combine networking and computing functions on a single chip to accelerate data center infrastructure. They’re designed to offload specific tasks from the CPU, improving overall data center efficiency and performance.

The ability to use NVIDIA BlueField-3’s compute power for multiple operations makes it compelling to use in conjunction with data management software like the WEKA® Data Platform. So, we teamed up with NVIDIA to test WEKA running on BlueField-3 DPUs.

WEKA provides a software solution that is hardware and cloud-agnostic. It’s composed of multiple software containers and processes that perform operations such as high-performance data storage, metadata distribution, distributed data protection, and more. WEKA utilizes technologies and algorithms that accelerate and secure IOs without specialized dedicated hardware.

WEKA currently runs on Intel X86 and AMD CPUs, but for our BlueField-3 tech preview, we wanted to leverage our upcoming ARM CPU support. Multiple organizations are looking to ARM CPUs for their lower price and power requirements. Public cloud vendors are in the process of releasing ARM instances and the new NVIDIA Grace CPUs are also based on an ARM architecture.

For our evaluation, it was a simple effort to deploy the WEKA ARM tech preview and test it on NVIDIA BlueField-3 DPUs. In this initial lab test, we wanted to validate running the WEKA client on NVIDIA BlueField-3 DPU cores while connecting to WEKA backend processes running on NVIDIA GPU servers.

The Test Environment

For the testing, NVIDIA provided a single GPU server in the NVIDIA lab, to which we remotely connected.

Diagram 1 – Test environment

As shown in diagram 1, we deployed multiple WEKA containers in converged mode on a GPU server, with each WEKA container using a single NVMe drive to simulate a complete WEKA cluster.

The WEKA client was successfully deployed on the BlueField-3 OS, eliminating the need to use any of the server’s resources and reserving them for any software that needs the CPU and RAM on the server itself. We then proceeded to create a WEKA filesystem and mount it on the WEKA DPU Client to validate IO functionality.

Diagram 2 – WEKA Client component running on DPUs

Diagram 2 shows the WEKA containers running on the single server (e155j-r750-02) while the WEKA client is running on the NVIDIA BlueField-3 DPUs (dorage-02-bf3).

This lab testing demonstrated the ability to run an ARM tech preview of the WEKA Data Platform on NVIDIA BlueField-3 DPUs. This deployment mode can be relevant for components of WEKA that enable multi-protocol support and physical segregation on the DPUs for secure environments that require access to the WEKA filesystem with complete CPU/Memory.

Next steps
WEKA Data Platform support for ARM is not yet a production feature but will be supported in a future release. When released, it will support NVIDIA BlueField-3 DPUs and other ARM environments such as public cloud ARM instances and the NVIDIA Grace CPUs.

Explore how the WEKA Data Platform works