~/talk

Andromeda: Performance, Isolation, and Velocity at Scale in Cloud Network Virtualization

Talkat NSDI '18 · 2018-04-09..04-11· [observability]
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
name:Jing Ai
links:none on file
name:Jon Olson
links:none on file
links:none on file
links:none on file
name:Nan Hua
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
name:Uday Naik
links:none on file
linkedin:vahdat
slides:none on file

Andromeda is Google Cloud’s network virtualization stack that uses a hierarchy of controllers plus Hoverboard gateways to avoid programming the long tail of mostly idle flows and offloads high‑bandwidth VM pairs once they exceed ~100 kb/s (results: >99.5% of traffic bypasses hoverboards). On hosts an OS bypass fast path plus per‑VM coprocessors handles packets; a single core hits ~38 Gbps and round‑trip latency ranges ~25 μs (15 μs w/ busy polling, ~9 μs with DPDK). Design scales to 100,000 VMs per network; 40k‑VM programming time falls from ~67s to ~2s; upgrades cause ~270 ms median blackout.

AI-written summary · about

  1. 0:00Intro
  2. 1:20Network Virtualization
  3. 2:18Andromeda Architecture
  4. 6:02Scaling with Hoverboards
  5. 6:40Hoverboard Offloading
  6. 9:24Andromeda Data Plane
  7. 11:28Data Plane - Fast Path
  8. 13:09Data Plane - Coprocessor Path
  9. 16:33CPU Efficiency Evolution
  10. 18:56Dataplane Hitless Upgrade 1/3
~/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.