April 25-29, 2016

Event Details

Please note: All times listed below are in Central Time Zone

Dive into Nova Scheduler Performance - Where Is the Bottleneck?

There are always rumors around nova scheduler. Like "Scheduler can handle at most 1k nodes without problems" or "I don't believe scheduler can handle so much requests using only one tiny process". Some would prefer to launch couples of scheduler instances to multiply the schedule throughput, but still others might worry about race conditions and decision conflicts that could outweigh the benefits. Cloud users are interested in the deployment strategies to maximize the schedule performance, and developers are curious about the mystery of race condition and its performance impact. Guesses and speculations are everywhere without evidences. So let's do some experiments and let data do the talk!

What can I expect to learn?
  1. The capability of one scheduler instance in handling concurrent requests with various compute nodes.
  2. The performance improvement using multiple scheduler processes.
  3. The scheduler race condition problems and its overhead.
  4. The suggested strategy to deploy OpenStack with schedulers.
Thursday, April 28, 3:10pm-3:50pm (8:10pm - 8:50pm UTC)
Difficulty Level: Beginner
  Yingxin Cheng has been working in Intel Open Source Technology Center since 2015. His major area is performance analysis and optimization in distributed systems. Specifically, he implemented a distributed-tracing solution in OpenStack and did a series of cluster behavior profiling at the scale of 1000 nodes. These works were published in OpenStack Summits (Austin and Barcelona) and FULL PROFILE
Software engineer in Intel FULL PROFILE
Intel Corp.
Ding JF is software engineer manager at Intel FULL PROFILE