Plan-Based Process Scheduling on HPC Nodes

worked on by: Kelvin Glaß

Outline

OUTLINE

Thesis Requirements

• Implement a prototype of a plan based scheduler in the current linux kernel version
• Analyse conditions of the plan (how many idle times, how are they distributed)

Weekly Status

Week 1 (CW 45)

Activities

• Research on the architecture of current linux versions
• Create a development infrastructure
• Start implementation of the scheduler

Results

• Created Qemu-Image, Scripts that build and execute the kernel
• Implemented prototype scheduler that switches after a certain time from PB-Scheduler to CFS
• Implemented linux module that starts the plan execution

Next Steps

• Enhance the scheduler so that the switch back from CFS to PB-Scheduler is in time
• Create a first table of contents
• Modify the kernel & module so that the plan can be set by the module

Problems

Week 2 (CW 46)

Activities

• Enhance prototype
• Enhance development infrastructure
• Create first table of contents

Results

• Working prototype scheduler that switches from PB-Sched. → CFS/Idle and Idle/CFS → PB-Sched.
• The plan can be set by a module
• It is possible to set a new plan after the first is execute
• Created infrastructure script, that converts a plan in csv → linux module + Makefile

Next Steps

• Start with theoretical work (e.g. find a definition of an unstable system)
• Create script that executes all modules one after another and aggregates the results

Problems

• Found a race-condition (which led to sporadic kernel-panic) in the prototype, caused by a missing put_prev_task call in the pick_next_task method of the PB-Scheduler

Week 3 (CW 47)

Activities

• Enhance test infrastructure
• Start writing first sections
• Start with analysis of criteria of a plan guaranteeing a stable system

Results

• Test script that executes a set of modules and aggregates the results
• First sections: Requirements of the PB-Scheduler prototype, Explanation of the abstract design of the Linux scheduler core functionality, Explanation of the implementation of the Linux scheduler (started)
• Divided stability criteria of a plan into sub problem: 1.: "What is the maximal task length without causing problems for other applications?", 2.: "What is the minimal idle time that allows to run processes from (1.) and run processes depending on the plan?"

Next Steps

• Continue with the definition of criteria
• Continue writing the Explanation of the implementation of the Linux scheduler
• Start analysis of different Queue states (CFS-Q full/empty, PB-Q full/empty)
• Start writing to explain how the PB-Scheduler prototype is implemented in Linux

Problems

• Running all tests showed that the delta between switching from idle to PB-Scheduler and switching from cfs to PB-Scheduler is still big. This could be fixed

Week 4 (CW 48)

Activities

• Write first draft of chapter "Plan Based Linux Scheduler", results in the "Evaluation" chapter, assumptions in the "Plan" chapter
• Install modified kernel on real hardware and execute all tests

Results

• First (!) draft of the "Plan Based Linux Chapter" (containing: Requirements, Why Linux, Architecture of current Linux scheduler, Module handling)
• Installed Lubuntu (16.10 with a 4.8.0-59-generic) on real hardware, installed modified kernel. The system ran stable with the new kernel (even the X environment).
• Executed tests to check whether the OS itself has a task execution time limitations. The result suggest that no limitation exists. Tested execution times of 1min, 20min, 60min.

Next Steps

• improve drafts
• Create test to determine the minimal free time of a plan (Approach: measure kernel thread exec. time during free time)

Week 5 (CW 49)

Activities

• Enhance kernel: Measure kernel thread execution time in a certain period
• Start tests with real HPC application to measure the amount of kernel thread execution time
• Write first draw about the enhancement and the measurement results

Results

• Enhanced kernel: Implemented measurement in __schedule() and required variables in the PB-runqueue structure
• Wrote first results and description of the enhancement
• Measured times for Matrix-Mul. and Primer-Nr.-Generation with MPI (100 test runs per input)

Next Steps

• Find a more representative app, define a testset, measure it, document environment, document results
• If possible: Derive abstract result from the measurement results.

Problems

• Find a representative example HPC application
• Running such an application is not easy in QEMU. Therefore it was necessary to run on real hardware.
• The first measurements results were wrong, because the idle thread is also a kernel thread, but the idle time should not be a part of the kernel thread execution time.

Week 6 (CW 1)

Activities

• Write first version of the evaluation
• Research for the first chapter that describes the plan creation, grid computing and the VRM

Results

• Write first version of the evaluation
• First sketches of the grid computing explanation

Next Steps

• Clean the second chapter so that the first chapter will the remaining main task

Week 7 (CW 2)

Activities

• Complete the code: rename variables, change indentation so that the code fits into LaTeX code boxes, Test the changed code
• Create State Transition Diagram and integrate into tex
• Correct result charts (error bars are < 0)
• Add explanation of the task_struct
• Add short explanation of each Linux scheduler module
• Check assessment criteria

Results

• Cleaned code
• Cleaner version of the seconds chapter

Next Steps

• Research for chapter 1
• Write chapter 1

Problems

As of Week 8 (CW 3)

• Iterative writing process