From jan.kofron at d3s.mff.cuni.cz Wed May 30 08:54:52 2018 From: jan.kofron at d3s.mff.cuni.cz (Jan Kofron) Date: Wed, 30 May 2018 08:54:52 +0200 Subject: Todays seminar in S7 Message-ID: Dear all, the seminar today takes place in S7 in five minutes! Looking forward! H. ========== Martin Děcký (Huawei Technologies): Data-Centric Computing In recent decades, the majority of mainstream computer systems we designed as CPU-centric and memory-centric. All the data that the computer processed always passed through the main memory in one way or the other and the code running on the central processor(s) manipulated the data. DMA controllers, co-processors, GPUs and intelligent NICs were used mostly as slave peripherals, although with some degree of computational acceleration and off-loading in recent years. One of the reasons for keeping the same basic architecture despite its drawbacks such as the ever-widening memory barrier (i.e. the difference between the latency of the CPU and the RAM) was arguably its conceptual simplicity. However, trends for a major change in the status quo are emerging into the mainstream. There are new non-volatile memory technologies that promise to marry the benefits of byte-addressable volatile memory and block-addressable permanent storage. Approaches such as data-centric computing and near-data processing propose to turn an ordinary computer into a distributed system where the distinction between CPU/RAM and peripherals is much more blurry than today, with the possibility to autonomously compute within and interchange data between all the hardware components. This talk presents a brief overview of the concepts and discusses consequences for the design of the software running on top of this new hardware.