Integrating Composable Memory Solutions into Storage intensive workloads

"Seema Mehta (Product Management) - Ampere Computing
Kapil Sethi (New Business Planning) - Samsung
Ravi Kiran Gummaluri ( ) - Micron

As memory-intensive applications become popular in high-performance computing and cloud environments- data centers face key challenges in achieving superior performance and lower TCO. Memory demands are growing- especially for workloads that rely heavily on in-memory data processing such as SPARK. Techniques like 2DPC to increase DRAM density or adding more CPUs and/or 3D stacking come at the expense of higher TCO- signal integrity issues and reduced bandwidth.. These challenges can be addressed by implementing either of the 2 approaches: CXL memory expansion or CXL memory pooling. Memory expansion using CXL attach-memory has already been discussed here. A composable memory system using the memory pooling approach is analyzed for SPARK and discussed in this paper. The analysis focuses on potential savings from unstranding memory resources in spark clusters with multiple nodes sharing memory to reduce the overall node count or memory footprint."