Siirry sisältöön

Tampere University: Low latency edge-cloud offloading software stack

Tampere University is Finland’s second-largest university with seven faculties conducting scientific research in technology, health and society providing the highest education within these fields. Customized Parallel Computing (CPC) is a research group in Tampere University. CPC’s main research focus is on design and programming methodologies of customized parallel computing platforms and real time implementations of challenging algorithms.

In edge compute offloading, a key challenge is often the latency. The type of workloads that can be outsourced from the local device down to a server residing in a near or far cloud is strictly limited by whether the overhead imposed by the communication chain and the API stack can be tolerated by the application.

Another major challenge is the heterogeneity of the computational resources; in scenarios where the edge device, such as a personal smartphone, a drone roaming around in a factory, or an intelligent car, is in motion, it comes in proximity with various types of edge servers with diversity of hardware (CPUs, GPUs, FPGAs) from various vendors. All of them should be programmable with the same unbiased “hardware abstraction layer” to enable fair competition between the processor suppliers.

Enabling novel edge-offloading scenarios

Within AISA, CPC works on enabling novel edge-offloading scenarios with various computational demands by developing a new open standard based software stack. The edge-offloading stack is based on the open OpenCL standard which is officially supported by the key CPU, GPU and FPGA vendors in the industry. Since the software architecture is a layered one, various productive programming models and formats, such as ONNX or SYCL, can be supported in a modular fashion on top of it.

However, even with the modern networking technologies such as the new 5G ones, offloading data-intensive applications across network is typically limited by the bandwidth. To address this, CPC also collaborates with Ultra Video Group in researching edge-offloading-suitable low-latency compression methodologies aimed for compute vision-only situations.

Learn more about Customized Parallel Computing group and CPC’s PoCL-Remote driver for edge-offloading using only the OpenCL API.

Learn more about other AISA use cases in Tampere University:
Ultra Video Group
Virtual-reality and Graphics Architectures Group

Contact person: 
Assoc. Prof. Pekka Jääskeläinen, pekka.jaaskelainen@tuni.fi