nova/doc/source/threading.rst
Joe Gordon a1e8fc6dd9 Update docs layout
* Explain these docs are for trunk (copied from ironic)
* All the docs in this repo are meant to be developer docs, so having a
  devref inside of the docs is redundant and just makes the docs more
  complicated to navigate. Move everything out of the devref folder and
  link to everything from main index.
* Move man pages into separate section. The man pages are pretty sparse
* right now, we should either make them useful or just delete them
* Remove dead docs from unused_docs list in doc/source/conf.py
* Shuffle docs landing page, move common referees to the top (API,
  hypervisor support matrix), Add a introduction section and more. The
  hope is the updated layout makes this document easier to navigate.
* Use maxdepth of 1
* Rename a few sections with what are hopefully better names

The next step is to prune out outdated documents and further cleanup
this page.

Change-Id: Iff453e47ccc902a0e72b1a5f6ce1ee939ff3a1a0
2015-05-15 10:47:18 -07:00

2.2 KiB

Threading model

All OpenStack services use green thread model of threading, implemented through using the Python eventlet and greenlet libraries.

Green threads use a cooperative model of threading: thread context switches can only occur when specific eventlet or greenlet library calls are made (e.g., sleep, certain I/O calls). From the operating system's point of view, each OpenStack service runs in a single thread.

The use of green threads reduces the likelihood of race conditions, but does not completely eliminate them. In some cases, you may need to use the @lockutils.synchronized(...) decorator to avoid races.

In addition, since there is only one operating system thread, a call that blocks that main thread will block the entire process.

Yielding the thread in long-running tasks

If a code path takes a long time to execute and does not contain any methods that trigger an eventlet context switch, the long-running thread will block any pending threads.

This scenario can be avoided by adding calls to the eventlet sleep method in the long-running code path. The sleep call will trigger a context switch if there are pending threads, and using an argument of 0 will avoid introducing delays in the case that there is only a single green thread:

from eventlet import greenthread
...
greenthread.sleep(0)

MySQL access and eventlet

Queries to the MySQL database will block the main thread of a service. This is because OpenStack services use an external C library for accessing the MySQL database. Since eventlet cannot use monkey-patching to intercept blocking calls in a C library, the resulting database query blocks the thread.

The Diablo release contained a thread-pooling implementation that did not block, but this implementation resulted in a bug and was removed.

See this mailing list thread for a discussion of this issue, including a discussion of the impact on performance.