Unraveling the mystery of NVIDIA's unified memory for safety-critical GPU systems

In the domain of safety-critical systems there is an increasing need for more compute-capable and higher performance devices. This comes from the dramatic increase on the software complexity caused by the newest intelligent and autonomous systems. Graphics Processing Units (GPUs), as multi-processing accelerators, are an ideal choice in this aspect due to their ability to handle big amount of data and computations. In order to ease the challenging task of programming such devices, vendors are continuously adding features, such as Unified Memory (UM), which allow the programmers to reduce their developing time on GPU applications. However, the use of GPU poses several challenges on safety-critical systems due to its close source nature and proprietary implementation. Therefore, this paper shows a deeper insight on how this feature works and present a way of exploiting this knowledge to reduce the execution time of applications using NVIDIA's UM feature. We demonstrate that these optimizations can make the data migrations predictable and reduce the required time.
This work was supported by the European Union’s Horizon Europe programme under the METASAT project (grant agreement 101082622). It was partially supported the Spanish Ministry of Economy and Competitiveness under grant IJC-2020-045931-I (Spanish State Research Agency / Agencia Espanola de Investigación (AEI) / http://dx.doi.org/10.13039/501100011033) and by the Basque Collaborative Research Grant Programme ELKARTEK under the AUTOTRUST project (KK-2023).
Peer Reviewed
Postprint (author's final draft)