Thymesis-P: An Approach to Rack-scale Disaggregation Over OpenCAPI
summary
Describes ThymesisFlow-P, a software-defined interconnect that uses OpenCAPI M1 and C1 modes to integrate off-tray components for rack-scale disaggregation. Explains a resource “stealing” model where servers request and use neighbor resources (notably disaggregated memory), and presents the design, implementation experience, and a live disaggregated memory demo.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:29A disaggregated memory example
- 3:44OK but, does it make sense to directly pipe system interconnect traffic over the network?
- 7:49Approach Outline
- 8:51M1 attached logic: Preparing OpenCAPI transactions for the network
- 10:38C1 attached logic support
- 11:34Remote Memory Hierarchy
- 12:58Bridging Ci with M1: RA- EA conversion and network ld handling
- 14:32High-level H/w design overview
- 16:20Designed to be a multi-brick to multi-network
- 17:35Circuit Network Logical Link Control support
- 20:01Network Link Failure Error Control
- 21:00Software Stack Overview
- 22:02Attaching remote memory: 2 different approaches
- 25:03Development approach
- 27:32401 Mhz timing closure
- 29:59Initial Performance Results (STREAM Benchmark)
- 32:33Bandwidth Aegregation datalink layer
- 32:54Bandwidth Aggregation LLC layer
related talks
45mBeyond Machine Boundaries: Experience with Memory Disaggregation
16mDisaggregation and the Application
31mUnshackling Memory!
16mOneSidedMW: Managing Disaggregated Memory Efficiently, Flexibly, and Securely with RNIC
21mChatty Tenants and the Cloud Network Sharing Problem
22mAn Open and Coherent Memory Centric Architecture Enabled by RISC V
16mPreview: Cloud Computing
23mBreaking the limits of x86
