Enhancement and optimization of new heterogeneous accelerators based on Cyborg
summary
Cyborg adds vGPU and NVMe SSD support, extends FPGA tooling (DPDK/programming and online program APIs) and implements online bind/unbind for accelerator cards; device profiles integrate with Nova and Placement so servers request composite accelerator pairs (GPU+FPGA) and Nova selects hosts by profile. Resource sync moves from per-change replication to data aggregation plus transactional writes into Placement resource tables, reducing synchronization time by more than 30%. Virtualization support increases GPU sharing and utilization, and the project ships API/agent/conductor changes so attachment, release and evacuation of accelerator devices work end-to-end.
AI-written summary · about
chapters (12)
- 0:00Intro
- 1:03Challenge of Heterogeneous Computing
- 2:56Present & Improvement
- 4:21Cyborg Architecture
- 6:20Nova Cyborg interaction
- 8:06Bind & Unbind
- 9:07VGPU management
- 10:30VGPU bind to VM
- 11:36Accelerator resources sync
- 13:55Enhance accelerator management
- 14:46Cloud platform architecture
- 16:09Smartnic Management & Plan
related talks
16mBOSH & GPU - a Match Made in Heaven
05mFrom Open Source to Real-World Scale
18mSmartNIC DPU Enablement in OpenStack with OVN
40mHow Can I Give Virtual GPU Resources to My End Users Seamlessly?
24mFPGAs in the Datacenter
05mDay 1 Keynote
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
47mGet off my GPU! Better Sharing of AI Accelerators in OpenStack with Blazar Reservations