Improving Data Communication Between SSD-NVMe And GPU
Indian researchers have conducted an in-depth study on different ways of directly communicating data stored on NVMe SSDs with GPU graphics processors. Its objective is to provide knowledge and guidance to computing platform designers to overcome the bottlenecks that occur when connecting these devices to the same bus in a computer, speeding up work in applications with high data consumption.
One of the keys to accelerating performance in computer systems is minimizing the barriers between the hard drives and the processor, something in which the standard NVMe interface and protocol has helped a lot. Using the PCI Express bus for direct communication between the CPU and SSDs.
A huge leap in performance is achieved, accelerating computing to a critical level for data-intensive applications. This approach has many possibilities, and one of them is to apply it to the data with which GPU graphics processors work, used in a large number of high-level mathematical and physical calculation applications, such as all those based on data science or processing. of three-dimensional images, among others.
This is of great interest to the industry, and two researchers from the Kanpur Institute of Technology in India have conducted an in-depth analysis of these technologies. Their goal was to find ways to combine NVMe SSD virtualization techniques with transfer mechanisms between NVMe-SSD and GPU.
In their article they explain that many modern GPU-compatible applications process large volumes of data that are hosted on secondary storage, and that there are already several proposals by other researchers to optimize the overhead of data transfer between devices connected to the same bus, for example, an NVMe SSD and a graphics card.
And they have dedicated their work to studying in depth the feasibility of different combinations of virtualization techniques of NVMe SSDs with data transfer mechanisms between these drives and a GPU. In addition, they have analyzed the impact of different data transfer parameters, and the main cause of performance in each case, quantifying the data transfer rate and the use of CPU resources.
Their analysis provides valuable information so that other researchers can delve into different approaches that will allow to overcome in the future the bottlenecks that occur in different SSD and GPU storage architectures connected to the PCI Express bus in computing platforms.
This, potentially, will allow the design of hardware and software architectures that increase the use of resources and accelerate the processing of stored data, using standards such as NVMe and technologies such as virtualization of high-speed SSD storage.