Transcript Linux Scheduling - School of Computing and Engineering

Linux Scheduling
Tanenbaum Ch. 10.3
Silberschatz Ch.21.5
cs431-cotter
1
Linux Scheduling Background
• Basic structure results from compliance with
POSIX standard
• 3 tiered system
• As a process ran, its priority decreased.
• When all processes were no longer
runnable, their priorities were recalculated.
• Approach did not scale well.
cs431-cotter
2
Linux Schedule Re-design
Objectives
•
•
•
•
•
Implement fully O(1) scheduling
Implement SMP scalability
Improve SMP affinity
Improve interactive performance
Optimize for 1-2 runnable processes, yet
scale well.
cs431-cotter
3
Kernel Scheduling vs. Process
Scheduling ???
• Kernel Scheduling
– Running process requests a system call
– Device driver, etc. generates a hardware interrupt
– Requires careful synchronization of data structure
access
– Kernel is preemptive (as of version 2.5)
• Process Scheduling
– Time-sharing algorithm designed for fair
preemptive scheduling among multiple processes
– Schedules for real-time tasks where absolute
priorities are more important than fairness.
cs431-cotter
4
Scheduling in Linux
Three classes of threads for scheduling purposes:



Real-time FIFO
Real-time round robin
Timesharing (for all non real-time processes)
cs431-cotter
Tanenbaum, Modern Operating Systems 3 e, (c) 2008 Prentice-Hall, Inc. All rights reserved. 0-13-6006639
5
Process Priorities
• Determined by priority and counter
• Static
– Used for real-time processes
– Priorities 0 - 99
• Dynamic
– Used for conventional processes
– Priorities 100-139
– Adjusted based on time run and priority class
cs431-cotter
6
Process Priority
• Based on original “nice” level of process.
– Default 0, range -20 to +19 (lower is better)
• I/O bound processes get higher priority.
• CPU bound processes get lower priority.
• Process characteristics adjusted
dynamically (range of ± 5)
• Highest runnable process is always
selected
cs431-cotter
7
Process Timeslice
• On creation, child process gets ½ of
parent process timeslice
• Higher priority also gets longer timeslice
– Low priority / less interactive ~ 10 ms
– Default ~ 150 ms
– High priority / more interactive ~ 300 ms
• When timeslice  0, process has expired
cs431-cotter
8
Runqueues
• Process scheduling implemented through
runqueues
– One per processor
– Includes array of lists for all priorities
• Runqueue struct:
–
–
–
–
# of runnable processes
Active priority array
Expired priority array
…
• Non-runnable processes that still have timeslice
and are blocked are managed in waitqueues
(waiting on whatever event is blocking them).
cs431-cotter
9
Scheduling in Linux
Figure 10-10. Illustration of Linux runqueue and priority arrays.
cs431-cotter
Tanenbaum, Modern Operating Systems 3 e, (c) 2008 Prentice-Hall, Inc. All rights reserved. 0-13-6006639
10
Load balancing in SMP
• Load balancing done to improve overall system
performance
– Called every 1 ms when processor is idle, otherwise
every 200 ms.
• If subject processor is more than 25% less than
busiest processor, processes are moved
– Prefer expired processes
– Prefer high priority processes
– Prefer “cache cold” processes
cs431-cotter
11
Scheduling System Calls
• nice ( ); (command line function)
•
•
•
•
sched_getscheduler( ); sched_setscheduler( );
sched_getparam( ); sched_setparam( );
sched_yield( );
sched_get_priority_min( );
sched_get_priority_max( );
• sched_getaffinity(); sched_setaffinity();
• sched_rr_get_interval();
cs431-cotter
12
Interrupt protection Levels
Top-half Interrupt Handlers
Highest priority
Bottom-half Interrupt Handlers
Kernel-system service routines
User-mode programs
cs431-cotter
Lowest Priority
13
Scheduler Performance
• Advantages
– Self-contained and easy to follow
– Adapts to a wide range of environments
– O(1) performance !!!
– Real-time performance good (although not
engineered for hard real-time systems)
– Optimized for I/O bound processes
(interactive users)
cs431-cotter
14
More Recent Schedulers
• Anticipatory Scheduler
– Linux Kernel Versions 2.6.0 through 2.6.18
– Optimized disk I/O by waiting a short time
after each read to see if another read from the
same file was coming.
– Big improvement for some services (Apache
Web server), but uneven effects elsewhere.
cs431-cotter
15
More Recent Schedulers
• Completely Fair Scheduler
– Linux kernel version 2.6.23 - ?
– Schedules processes based on which
process has used the least of its time
quantum.
– Stores processes in a tree (red-black)
structure to facilitate picking the lowest user.
– O(log N) performance.
cs431-cotter
16
Summary
• Kernel is preemptive.
• Linux supports 3 schedulers:
– FIFO for real-time processes
– RR for interactive, I/O bound real-time processes
– OTHER for regular processes
• User processes use a priority queue
mechanism.
– Process priority is dynamic based on base
priority and time remaining in time quantum.
Each priority class runs as RR
cs431-cotter
17
Questions
• How does the Linux scheduler decide which
process/thread to run when it schedules a new
process/thread for the CPU?
• How does Linux ensure that real-time processes
always run before any other processes?
• What timeslice values do real-time processes
use?
• How does Linux dynamically adjust the
scheduling for I/O intensive processes (giving
them a higher priority)?
• Once all runnable processes have exhausted
their timeslices, how does the scheduler
recalculate/ restart those processes?
cs431-cotter
18