Breaking the Energy, Cooling, and Resource Ceiling for Future Supercomputer Architectures Through the Use of Composable Disaggregated Infrastructure
Michael Aguilar, Catherine Appleby · 2024
SAND2024-01831CBreaking the Energy, Cooling, and Resource Ceiling for Future Supercomputer Architectures Through the Use of Composable Disaggregated Infrastructure 1. Quick overview of a Beowulf HPC Architecture, the majority of the HPC Systems 2. The 'Big Drawback' to current Beowulf HPC Architectures 3. Composable Disaggregated Infrastructure (CDI)---Software Derived Hardware Architectures 4. Design Considerations for a Composability Manager on a Large-Scale HPC System 5. Introducing Sunfish 6. Sunfish Core Services 7. Sunfish Hardware Agents 8.The Sunfish Composability Management Framework 9. How Machine Learning can help us allocate CDI Resources and Algorithm Design 10. Acknowledgements and Questions 2 Composable Disaggregated Infrastructure (CDI)---Software Derived Hardware Architectures 9 Mitigate Filesystem Cache Thrashing by dynamically adding memory to the IO node (eg.ZFS Adaptive Replacement Cache, XFS Buffer Cache, etc.) to prevent Virtual Memory Swaps from disk.Cache fills up Cache fills up Dynamic IO Node Augmentation of Memory Fabric Fabric