Project Task Page: Axon Quiescence Optimisation

Description

Modifying axon so that components are not scheduled (ie. run) when they are paused. If all components are paused, then the whole Axon system sleeps until woken - it goes quiescent. As it stands, Axon does not do this - the scheduler keeps all microprocesses in a single list and polls all of them continually.

This is also an "inversion of flow control" - since it enables an Axon based system to become reactive to events, rather than polling.

Benefits:

• General performance and responsiveness improvement - processor cycles will not be wasted on idle microprocesses
• Better cooperation with other processes on a computer - Axon based systems will no longer hog processor cycles, even when inactive

Needed to make practical high data throughput Kamaelia systems, such as KamaeliaMacro.

Inputs

Task Sponsor: Michael (MS)

Task Owner: Michael, Matt (MH)

Developers: Matt (initial sketches and final code dev), Michael (final merge)

Users: Michael, Matt

Interested third parties: none

Requirements

• MUST cause an Axon system to relinquish the cpu when idle
• MUST not change the behaviour of an Axon system from the perspective of individual components
• MUST be threadsafe
• SHOULD only schedule and run microprocesses that are actually active at a given time (scheduler should properly pause() microprocesses and also unpause() them)
• SHOULD not noticeably impact performance
• SHOULD be responsive to external stimuli when quiescent (eg. Selector component gets woken when data is available)
  

Relevant Influencing factors:

• Needed for Kamaelia Macro

Outputs

Expected

• Modifications to Axon (initially in branch, then merged into mainline code)
• Modifications and additions to Axon test suite
• Improved performance
  

Actual

Code:

Experiments (in order):

• /Sketches/MH/Inversion/miniaxon.py
• /Sketches/MH/Inversion/cleverscheduling.py
• /Sketches/MH/Inversion/aggregatedevents.py

CVS branch: private_MH_axon_flowcontrolinversion

Mainline code: /Code/Python/Axon/Axon/

• Modified: most/all source files
  

Mainline code: /Code/Python/Axon/Axon/Tests

• Tests removed that were now inapplicable
• New tests added

Bugfixes for components in mainline: (turned out several components did not use self.pause() properly)

• /Code/Python/Kamaelia/Kamaelia/Internet/Selector.py - made threaded and pausing
• /Code/Python/Kamaelia/Kamaelia/Internet/TCPServer.py
• /Code/Python/Kamaelia/Kamaelia/Chassis/ConnectedServer.py
• /Code/Python/Kamaelia/Examples/example2/SimpleStreamingSystem.py
• /Sketches/MPS/Macro/UnixPipe.py
• /Code/Python/Kamaelia/Kamaelia/File/*
• /Code/Python/Kamaelia/Kamaelia/Util/passThrough.py
• /Code/Python/Kamaelia/Kamaelia/Visualisation/PhysicsGraph/lines_to_tokenlists.py
• /Code/Python/Kamaelia/Kamaelia/UI/*
• /Code/Python/Kamaelia/Examples/example11/Ticker.py
• /Sketches/MH/Sketcher/*
• /Sketches/MPS/Paint/Paint.py
• /Sketches/MH/Editor2/Sandbox.py
• /Code/Python/Kamaelia/Kamaelia/Util/Introspector.py
• 
  

Realistic possibilities arising as a result of activity on this task

Axon's support for the following was removed; and needed to be added back in again:

• limited "pipewidths" - linkages that could only carry a limited number of items data before becoming 'full' - (a rate limiting mechanism)
  
• waking a component when a message leaves on of its outboxes (free space to send to)

Task Log

• 4 May 2006 : Added developer Matt
• 4 May 2006 : Task status changed to running.
• 4 May 2006 : Matt : worked on inital sketches/experiments
  
• 4 May 2006 : Matt : work commenced on private_MH_axon_flowcontrolinversion branch
• 7 May 2006 : Matt : finished sketches/experiments
• 14 May 2006 : Matt : began retrofitting new threading and pausing support to key Kamaelia components
  
• 31 May 2006 : Matt : completed development
• 31 May 2006 : Michael : Allocated to merge branch into mainline
• 1 June 2006 : Michael : Merge completed
• 1 June 2006 : Task status changed to completed
  
• 1 June 2006 : Removed developers, Michael, Matt
• 22 February 2007 : Matt : Project Task Page created
  

Discussion

This work mainly centred on modifications to the scheduler - replacing its main loop with on that wouldn't iterate through inactive microprocesses, and providing a thread safe mechanism for submitting requests to it to activate, pause and unpause microprocesses.

Two variants were experimented on in /Sketches/MH/Inversion. cleverscheduling.py used queues for the pause/unpause/activate requests. aggregatedevents.py attempted to improve efficiency by building a table of changes, rather than simply queuing repeat requests. The former approach was adopted as it was slightly simpler to implement.

The following components notably underwent substantial overhauls to take advantage of the new quiescence capability, and in some cases, fixing buggy uses of self.pause():

• Selector - made into a threaded component that blocked on select calls and would pause when idle
• Pipeline/Graphline/Carousel - improved handling for child shutdown on termination
• Pygame components - separated event polling into a threaded component to allow pausing when inactive; and modified existing components to explicitly send a request when they need to be redrawn, rather than assuming it happens continuously.
• Introspector - new method calls were added to the scheduler for it to use, so it could introspect without knowing too much about the scheduler's internal structure.

-- Matt Hammond, 22 February 2007