BROWSE
Detail
Conference Paper
Ethernet-Based Memory Expansion with Cache-coherent Architecture for Disaggregated Memory Systems
Cited 0 time in 
Share 
Share
- Authors
- Song-woo Sok, Young woo Kim, Hyuk Je Kwon, Jinmee Kim, Hag Young Kim, Kang Ho Kim, Seung-Jun Cha
- Issue Date
- 2024-12
- Citation
- International Conference on Big Data (Big Data) 2024, pp.8780-8782
- Language
- English
- Type
- Conference Paper
- Abstract
- Memory-centric computing is essential in big data analysis for minimizing data movement and latency, enabling efficient, real-time processing of large-scale datasets. We proposed MECA (Memory-Expansion Cache-coherent Architecture), a memory-centric design that leverages OmniXtend protocol over Ethernet to establish cache coherence across large-scale memory pools, utilizing RISC-V architecture to ensure efficient and consistent memory sharing. Performance evaluations show MECA achieves up to 98% of local memory performance for matrix-intensive workloads through hybrid memory optimization, significantly reducing latency and bandwidth limitations. These results highlight MECA's potential as an open-source alternative to proprietary solutions like CXL, promoting innovation and scalability in memory-centric computing for data-intensive applications.
- KSP Keywords
- Bandwidth limitation, Big data analysis, Cache coherence, Data intensive applications, Data movement, Ethernet-based, Hybrid memory, Large-scale datasets, Local Memory, Memory System, Memory optimization