Increasing DDR4 SDRAM throughput in parallel workloads
Authors:
- 1 March 2020
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-4
Abstract
DRAM chips and modules have been essential building blocks of electronic equipment ranging from embedded systems to supercomputers for decades. However, the chip technology trends have created a gap between processor and memory performance, aggravated by multi-core processor design. Not only memory bandwidth becomes insufficient in various applications, the actual throughput, or efficiency, of the memory is reduced in multi-threaded applications.In this work, we analyze the memory efficiency problem in parallel workloads, identify two approaches to solving it, survey and assess some of the possible solutions. In our experiments with DDR4 SDRAM controller, we achieve the average efficiency of 71 percent in 16-threaded workloads.Keywords
This publication has 9 references indexed in Scilit:
- Development of memory controller for today’s Elbrus microprocessorsRadio Industry (Russia), 2019
- An Exploration of Approaches to Instruction Pipeline Implementation for Cycle-Accurate Simulators of «Elbrus»Proceedings of the Institute for System Programming of the RAS, 2019
- CART: Cache Access Reordering Tree for Efficient Cache and Memory Accesses in GPUs2018
- MOCA2017
- A Survey of Recent Prefetching Techniques for Processor CachesACM Computing Surveys, 2016
- Adaptive Runtime-Assisted Block Prefetching on Chip-MultiprocessorsInternational Journal of Parallel Programming, 2016
- Power/Performance Trade-Offs in Real-Time SDRAM Command SchedulingIEEE Transactions on Computers, 2015
- BuMP: Bulk Memory Access Prediction and Streaming2014
- Guided region prefetchingACM SIGARCH Computer Architecture News, 2003