Octopus: Enhancing CXL Memory Pods via Sparse Topology
summary
Octopus constructs a sparse CXL topology that groups servers into islands (BIBD intra-island, expander inter-island) to trade overlap for pooling efficiency. A 3-server prototype and 96-server simulation (measured device specs, 1.5 m copper limits) show 65% of memory becomes poolable versus 35% with CXL switches, achieving ~16% less memory to provision and 3.2× faster RPCs than in-rack RDMA (2.4× faster than CXL switches), yielding about 3–5.4% net server cost savings.
AI-written summary · about
chapters (10)
- 0:00Octopus and CXL memory pods
- 1:42Memory cost crisis and CXL pooling
- 3:27CXL pod requirements and limits
- 6:02CXL switch and multi-port challenges
- 8:34Sparse topology and Octopus design
- 10:40Balancing pooling and communication
- 13:05Octopus hierarchical island topology
- 14:50Evaluating Octopus memory pooling
- 16:18Cost analysis and CAPEX impact
- 17:01Future research for CXL pods
related talks
17mDuhu: Shared Disaggregated Memory for Distributed Data Processing Frameworks
19mCluster Optimization: Harnessing DPUs for Ambient Mesh Workloads
25mGreenOps at Scale: The 50TB Memory Diet for a Petabyte Platform
01h 07mOpenCAPI BoF - A Memory Centric Fabric for a Data Centric World
20mRDMA at Cloud Scale: The OCI Experience
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
27mSustainable, Modern, and Flexible Solutions for Cloud Environments
14mHyperscale Infrastructure Q&A