Orca: Server-assisted Multicast for Datacenter Networks
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Orca claims a practical multicast primitive for datacenters that splits forwarding work between switches and servers, implements a controller plus per-rack Orca agent, and runs on a testbed and large simulation. It builds label formats that are fixed-size (about 19 bytes fixed label), pushes leaf label resolution to servers, and replicates tiny per-rack state for fault tolerance. Evaluation on a 27,648-host simulation with 1M sessions and a testbed shows reduces switch state by up to two orders of magnitude and zero redundant traffic.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:36Group Communication in Datacenters
- 1:48Multicast in Datacenter Networks Can benefit from multicast
- 3:03Challenges of Multicast Forwarding in DCNS Large-scale multi-tenant environmen
- 4:57Our Goal
- 7:30Our Approach
- 7:54Orca Labels
- 8:44Orca: Leaf Labels • Leaf switch forwards multicast packets to Orca Agent
- 9:15Orca: Source Label
- 10:30Orca Architecture
- 11:49Fault-tolerant Forwarding
- 12:53Evaluation using Simulations A Leaf-spine topology with 27K hosts
- 13:50Orca vs Elmo
- 15:46Evaluation in a Testbed
- 16:00Benefits of Orca to Applications
- 17:21Agent Scalability
- 17:47Conclusions Orca is an efficient multicast system for DCNS • The key idea is to move parts of forwarding to servers
related talks
10mLightning Talk: Introducing ORCA: Load Balancing With Endpoint Provided Load Data
16mRingLeader: Efficiently Offloading Intra-Server Orchestration to NICs
20mRDMA at Cloud Scale: The OCI Experience
30mOrion: A Distributed File System for Non-Volatile Main Memory and RDMA-Capable Networks
18mCloudcast: High-Throughput, Cost-Aware Overlay Multicast in the Cloud
01h 01mOrca: A Modular Query Optimizer Architecture for VMware Greenplum
21mLarry: Practical Network Reconfigurability in the Data Center
22mExpanding across time to deliver bandwidth efficiency and low latency
