Scheduling Meeting 12601

1
Scheduling Meeting (1-26-01)

• Status
• Work in progress
• Preliminary ideas on continuum scheduler

2
Status

• Past 2 weeks (completed)
• Prepared graphs for Randy. Fixed apps that were
  not run in awhile.
• Need to setup regression testing
• Rewrote mpeg_encode to eliminate dynamic
  composition
• saves shared memory
• can be used with static scheduler
• found several bugs in tdfc backend, reported to
  Eylon.
• Investigated memory leak in Dynamic Scheduler
• it is sufficiently small to ignore for now
• Rewrote stream buffer allocation code to save
  shared memory and permit running large apps (e.g.
  mpeg).

3
Status

• In progress
• Early Stall Detection -- completed, has
  inaccurate detection -- needs rates to proper
  compute critical set
• need better treatment for CPU
• Investigating how to fully isolate the effects of
  user application on the scheduling
• Several user apps (jpeg, mpeg, IIR, wavelet) are
  ready to setup automatic regression testing and
  profiling.

4
Scheduling -- need to do

• Rates
• small amount of changes
• all the infrastructure is in place
• Custom size buffer allocation
• modify resource allocation code
• various optimization criteria are possible,
  balance CP utilization and amount of data
  transfer
• other modules are in place (sequence, script
  generation), assuming no feedback

5
Continuum Scheduling

• Propose
• fully event-driven script scheduler
• Hardware support
• each individual compute node will queue events
  (interrupts) to the CPU
• interrupt request contains an argument, interrupt
  vector number (address in the scheduling script
  where handle routin is).

6
Continuum Scheduling

• Software/Event Handlers
• Two types of operations are needed
• Load handler
• Dump handler
• Each CP/CMB is configured to point to its dump
  handler.
• When dump condition occurs (stall, empty/full
  buffer, done) a CP/CMB issues an event pointing
  to an appropriate dump handler.

7
Continuum Scheduling

• Software/Event Handlers
• global array controller
• configured to detect a combination of dump
  events,
• issue load event(s) referring to appropriate
  pages/segments.
• Since all events are queued up for the handler
  the order will be preserved
• first all required dumps, then loads.

8
Continuum Scheduling

• Questions
• how to compute dump conditions
• CP can not issue its own dump events (except in
  case of done)
• similar problem to computing a set of critical
  memory segments in case of early stall detect
• can token availability be reintroduced?
• Independent load and dump event handlers
  permit skipping nodes in a dependency graph.
• with exception of CMB segfaults, can script be
  loaded in array controller to relieve CPU from
  doing any scheduling work?