~/talk

Flor: An Open High Performance RDMA Framework Over Heterogeneous RNICs

on YouTube 15m
Talkat OSDI 2023 · 2023-07-10..07-12· [networking]

speakers

name:Qiang Li
links:none on file
name:Yixiao Gao
links:none on file
name:Xiaoliang Wang
links:none on file
name:Haonan Qiu
links:none on file
name:Yanfang Le
links:none on file
name:Derui Liu
links:none on file
name:Qiao Xiang
links:none on file
name:Fei Feng
links:none on file
name:Peng Zhang
links:none on file
name:Bo Li
links:none on file
name:Jianbo Dong
links:none on file
name:Lingbo Tang
links:none on file
name:Hongqiang Harry Liu
linkedin:hongqiang-liu-05b0b727
name:Shaozong Liu
links:none on file
name:Weijie Li
links:none on file
name:Rui Miao
links:none on file
name:Yaohui Wu
links:none on file
name:Zhiwu Wu
links:none on file
name:Chao Han
linkedin:chaohandata
name:Lei Yan
links:none on file
name:Zheng Cao
links:none on file
name:Zhongjie Wu
links:none on file
name:Chen Tian
links:none on file
name:Guihai Chen
links:none on file
name:Dennis Cai
linkedin:dennis-cai-45699a2b
name:Jinbo Wu
links:none on file
name:Jiaji Zhu
links:none on file

resources

video:https://youtu.be/cmlTZr0yF8s
slides:none on file

summary

Flor is an RDMA framework that keeps the RoCEv2 data path in RNIC hardware but moves vendor-specific control logic into software running on NPUs/DPUs/host CPUs, implements dynamic chunking, selective UC-first transport, and software reliability/congestion modules. Implemented on Intel E180, Mellanox CX-4/CX-5, Broadcom RNICs, it matches vanilla RDMA performance under stress tests including 1/4096 packet loss, 6000:1 incast, and reduces CX-4/CX-5 bandwidth gap from 24.3% to 1.3% in production.

AI-written summary · about

chapters (14)

  1. 0:00Flor RDMA framework
  2. 0:29RDMA with heterogeneous RNICs
  3. 1:34Performance gaps in RNICs
  4. 2:16New RDMA production needs
  5. 3:22Limitations of RDMA frameworks
  6. 4:03Flor's design key findings
  7. 5:28Flor architecture and control path
  8. 6:52Dynamic chunking for RDMA
  9. 8:39Unreliable Connection transport
  10. 9:42RTT and congestion control
  11. 10:29Evaluation setup and throughput
  12. 11:37Large-scale deployment results
  13. 12:45Alibaba application performance
  14. 13:29Conclusion and future work

related talks

~/search
type to search 6,249 conferences, 138,987 speakers and 185,172 talks
~/help
cmdctrl k
search every conference, speaker and series
/
focus the filter search
h
home
c
conferences
s
speakers
e
series
t
light / dark — the status bar button does the same
?
this help
esc
close this, then clear every filter — or just the search, while you are typing in it

Ignored while you are typing. Pages without a filter rail have no search to focus.