This document discusses semantical cognitive scheduling, which aims to improve process scheduling by considering the current state of the operating system. It proposes assigning a utility value to each process-state pair based on the semantical progress gained from running that process in that state. This allows the scheduler to maximize total utility by considering the current system state. Obtaining utility values can be done through machine learning, user feedback, or from software vendors. While the general problem of finding the optimal scheduling order is hard, complexity can be reduced by assuming the number of states is bounded. Dynamic programming and greedy approaches are proposed to find high-utility schedules in polynomial time by exploiting this bounded state space assumption.

1. Semantical Cognitive SchedulingIgal ShilmanBen-Gurion University, Be’er-ShevaJoint work with: Prof. ShlomiDolevBen-Gurion University, Be’er-ShevaDr. AviMendelsonMicrosoft Research, Hertzlia

2. AgendaBackground & MotivationOur Approach Obtaining the utility valueSystem SettingsProblem ComplexityBounded state space

3. Background Cont.A processes is a program in execution.The Operating System (OS) manages the processes.

4. The Process SchedulerA major component of operating systems. Scheduling policies:user interactivitythroughputreal time responsiveness

5. User interactivityMeasured as the amount of time passed from the user’s action and the program response.Two Major algorithms that adhere to user interactivity are:Priority basedGuaranteed based

6. MotivationSuppose we have Skype running but no internet connection present.Observe the following behavioral pattern:Tries to connect(I/O)Tries to connect(I/O)…Connection Failed.Wait a whileConnection Failed.Wait a whileTime

7. Motivation Cont.Clearly semantical progress that Skype does (and therefore benefit to the user) is: None!AgendaBackground & MotivationOur Approach Obtaining the utility valueSystem SettingsProblem ComplexityBounded state space

8. Our ApproachProcess behavior depends greatly on the state of the OS. A good scheduler will take into account the current state of the system. There is nosuch scheduler!

9. Our Approach – Def.A state is an abstract description of the OS’s memory.The utility valueassigns a value to the semantical progress gained when some process is scheduled in a certain state.Utility valuestateNo Internet connection presentConnection is present

10. Our Approach – Def .A state is an abstract description of the OS’s memory.The utility valueassigns a value to the semantical progress gained when some process is scheduled in a certain state.Stateis the position of the RW headP want to get block yyP’ want to get block xx

11. Our Approach – Def.A state is an abstract description of the OS’s memory.The utility valueassigns a value to the semantical progress gained when some process is scheduled in a certain state.State Utility valueCoefficients Reflects importance

12. AgendaBackground & MotivationOur Approach Obtaining the utility valueSystem SettingsProblem ComplexityBounded state space

13. Obtaining the Utility ValueClassical scheduling techniques assume no (a priori) knowledge.The amount of unique classes of applications that the scheduler meets is relatively small.We propose few approaches for obtaining the utility value.

14. Obtaining the Utility ValueBlack box approach.Use machine learning theory.Correlate state and application compatibility statistically.Learning users feedback.Interpret user actions (like clicks, amount of time that the window reminded focused, etc)

15. Obtaining the Utility ValueWhite box approach. Provided by software vendors.An agent that lives inside the process and communicates with the OS-defined APIVendors interests.

16. AgendaBackground & MotivationOur Approach Obtaining the utility valueSystem SettingsProblem ComplexityBounded state spaceThe use of the dependency graph

17. SystemSettingsA set of n processes,A set of r states, A utility functionwhich assigns a value to each process and state.A transition function.An initial stateWe wish to find an ordering which will maximize:

18. Example

19. Example

20. AgendaBackground & MotivationOur Approach Obtaining the utility valueSystem SettingsProblem ComplexityBounded state space

21. Problem ComplexityTheorem 1: Any algorithm that solves the utility maximization problem, for which , has to consider any ordering, thus its time complexity is . is a black-box. Need to explore Unique edges.Each edge introduce new Weight.

22. Problem ComplexityTheorem 1: Any algorithm that solves the utility maximization problem, for which , has to consider any ordering, thus its time complexity is .This problem is HARD!

23. AgendaBackground & MotivationOur Approach Obtaining the utility valueSystem SettingsProblem ComplexityBounded state space

24. Bounded state spaceGeneral case: We assume, that the number of abstract states is bounded, in fact: ≤For example ACPI has a constant (140) number of states.

25. Dynamic programming approachWe need to scheduleInitial state: sRecursive equation:The Problem is reduced to finding an optimal ordering of processes which start at:

26. Dynamic programming - Cont.There are subsets of P.Each subset has to be considered with any of the r states from S. Number of recursive calls is Running time is: Space is: Which is an improvement over

27. Non-optimal approachGreedy approach:Chose the process that introduces the maximal utility value in the current state.The batch could be scheduled in p1p1p1S3S2S1

28. Non-optimal approachGreedy approach:Chose the process that introduces the maximal utility value in the current state. The batch could be scheduled in The method could be extended naturally for a parameter t:EvaluateTime: O(nt) (Bounded depth DFS)