Supporting Massive DLRM Inference through Software Defined Memory
Ehsan K. Ardestani, Changkyu Kim, Seung Jae Lee, Luoshang Pan, Jens Axboe, Valmiki Rampersad, Banit Agrawal, Fuxun Yu, Ansha Yu, Trung Bao. Le, Hector Yuen, Dheevatsa Mudigere, Shishir Juluri, Akshat Nanda, Manoj Wodekar, Krishnakumar Nair, Maxim Naumov, Chris Petersen, Mikhail Smelyanskiy, Vijay Rao · 2022 IEEE 42nd International Conference on Distributed Computing Systems (ICDCS) · 2022
Deep Learning Recommendation Models (DLRM) are widespread, account for a considerable data center footprint, and grow by more than 1.5x per year. With model size soon to be in terabytes range, leveraging Storage Class Memory (SCM) for inference enables lower power consumption. This paper evaluates the major challenges in extending the memory hierarchy to SCM for DLRM, and presents different techniques to improve performance through a Software Defined Memory. We show how underlying technologies such as Nand Flash and 3DXP differentiate, and relate to real world scenarios, enabling from 5% to 29% power savings.