Reimaging the Cloud: Data-Centric Multi-Acceleration at Scale

Zoom link for remote attendees: password 123123

Abstract: In the post-Moore era, scaling systems out and specializing in compute aim to propel the age of planet-scale applications. However, the memory wall, coupled with datacenter system taxes, has created a Data Motion bottleneck. This bottleneck encompasses the overhead of both moving data and preparing it for accelerated computation. This talk first discusses the initial steps toward a data-centric view, focusing on near-data processing innovations that, besides accelerating computation, also address data motion and datacenter taxes. Then, we discuss the reality that real-world applications are not just machine learning kernels but comprise kernels across different domains. Thus, to unleash the next wave of computing, we need a shift toward data-centric, cross-domain multi acceleration. In contrast to the status quo, which mostly uses specialization for compute, we present data motion acceleration for chaining cross-domain multi-accelerators. These innovations set the stage for rethinking the data delivery hierarchy, such that it integrates data motion accelerators  within memory, storage, and interconnection as first-class primitives. This re-envisioning is the focus of my future work, which aims to build the next-generation datacenter that accelerates planet-scale, cross-domain applications.

Bio: Mohammad Alian is an Assistant Professor in the EECS Department at the University of Kansas. His research seeks to reimagine data delivery in the next generation datacenters to unleash the full potential of compute acceleration in scale-out systems. Mohammad obtained his Ph.D. from the ECE Department at the University of Illinois Urbana-Champaign in 2020. His research led to the first industry memory module, Samsung AxDIMM, that integrates near-data processing. His research works were runner-ups for the Best Paper award in HPCA 2017 and MICRO 2018. His work has also received an honorable mention in IEEE MICRO Top Picks. He is currently a task-lead PI and the broadening participation champion in the DARPA/SRC ACE Center for Evolvable Computing. Mohammad is the recipient of the NSF CAREER award, and his team achieved second place in the Samsung Open Innovation contest in 2022. More information is available at https://alian-eecs.ku.edu