~/talk

SelfTune: Tuning Cluster Managers

Talkat NSDI 2023 · 2023-04-17..04-19· [machine-learning, devops-and-delivery]
links:none on file
links:none on file
name:Lei Zhao
links:none on file
links:none on file
links:none on file
links:none on file
links:none on file
slides:none on file

SelfTune is a lightweight API plus learning engine that plugs into cluster managers via simple calls like predict and set_reward, uses a continuous bandits algorithm, and tunes real and discrete parameters online with low sample complexity. Deployed to tens of thousands of machines for Microsoft’s background scheduler, it continuously adjusts a concurrency knob to raise throughput up to 20%. In benchmarks it tunes ~100 parameters for DeadStar, cuts rounds-to-good-config from ~30–40 down to about 8, and lowers iteration cost.

AI-written summary · about

  1. 0:00Intro
  2. 0:24Cluster management frameworks
  3. 4:30Configuring cluster managers
  4. 7:54Key Features
  5. 9:08Environments, rewards, optimal configs vary with time
  6. 9:59Evaluation: Three cluster management settings
  7. 13:13Azure Function-as-a-Service
  8. 14:25Hybrid Histogram Policy
  9. 16:14Summary
~/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.