Application Performance in the QLinux Multimedia Operating System

1
Application Performance in the QLinux Multimedia
Operating System

• Jun Wang

2
Overview

• H-SFQ CPU/Packet Scheduler
• Cello Disk Scheduler
• Lazy Receiver Processing (LRP)
• Related Works
• Open Discussions

3
H-SFQ CPU/Packet Scheduler
4
Requirements for New CPU Scheduler
Example 1 Consider a MPEG decoder and some
conventional processes running together. The CPU
scheduler is the standard time sharing
scheduler. The MPEG decoder has a soft real-time
demand which is 25 frames per seconds. When the
number of conventional processes increases, what
is the problem?
5
Requirements for New CPU Scheduler
Example 2 Consider some conventional processes
and some hard real-time processes such as control
system of aircraft running together in a
real-time OS. The CPU scheduler is Earliest
Deadline First (EDF). When the number of hard
real-time processes increases, what is the
problem?
6
Requirements for New CPU Scheduler

• Problem Analysis
• A conventional CPU scheduler can not adapt
  multiple classes of applications simultaneously.
• A conventional CPU scheduler can not isolate
  different classes of applications so that
  increasing the load of one class of applications
  does not influence other classes of applications.
  

7
H-SFQ CPU Scheduler A Sample
Hierarchy
8
H-SFQ CPU Scheduler
Features

• Multiple level scheduling structure
• ?class-independent scheduler for non-leaf node
  class-specific scheduler for leaf node
• Support for multiple service classes
• ?solving the simultaneous occurrence problem
• Predictable resource allocation
• ?solving the isolation problem
• Proper accounting of resource usage
• ?accounting of CPU usage by virtual time

9
SFQ CPU Scheduler Algorithm
Description(1)
Example Consider two threads A and B with
weights 1 and 2,respectively,which become
runnable at time t0. Thread A will run 80ms to
finish, and Thread B will run 110ms to finish.
But both of them will block at some time. Let the
scheduling quantum for each thread be 10ms.
10
SFQ CPU Scheduler Algorithm
Description(2)
11
SFQ CPU Scheduler Properties

• SFQ achieves fair allocation of CPU regardless of
  variation in available processing bandwidth
• SFQ does not require the length of the quantum to
  be known a priori
• SFQ provides bounds on maximum delay incurred and
  minimum throughput achieved by the threads

12
H-SFQ CPU Scheduler
implementation
Weight Start-tag Finish-tag
13
H-SFQ CPU Scheduler
Evaluation(1)
14
H-SFQ CPU Scheduler
Evaluation(2)
15
Open Discussions Section One

• How to convert different user requirements into
  weights in H-SFQ?
• HLS another hierarchical framework for soft
  Real-time schedulers
• ?Multiple kinds of schedulers in non-leaf
  nodes
• ?Support both relative allocations and
  absolute allocations

16
Cello Disk Scheduler
17
Requirements for New Disk Scheduler
Conventional Disk schedulers experience the same
simultaneous occurrence problems and isolation
problem as the conventional CPU schedulers
18
Cello Disk Scheduler
Architectural Principles
19
Cello Disk Scheduler Features

• Multiple level scheduling structure
• Class-independent scheduler combines weight and
  disk service time into its decision of
  scheduling.
• Class-specific scheduler steal the slack time of
  soft real-time requests for interactive requests
  as much as possible

20
Cello Disk Scheduler
Implementation Ideas(1)
Example There are 3 classes disk requests with
equal weights in the system, one is soft
real-time class with 20ms deadline for every
request and 6ms service time, one is interactive
class with 3 ms service time, the last is
throughput-intensive class. The following figure
shows the contents of pending queues and
scheduled queue at the beginning of a certain
time interval which is 45ms.
soft real-time r3 r2 r1
interactive
Scheduled Queue i3 i2 i1
throughput-intensive
t1
21
Cello Disk Scheduler
Implementation Ideas(2)


Switching from one class to another class when
one of three conditions is met ?the pending
queue for a class is empty ?the sum of requests
outweigh the classs proportion ?the sum of
requests outweigh the current time interval,
soft real-time r3 r2
interactive
Scheduled Queue i3 i2 i1
r1
throughput-intensive t1
soft real-time r3
interactive
Scheduled Queue i3 i2 i1
r2 r1
throughput-intensive t1
22
Cello Disk Scheduler
Implementation Ideas(3)
Slack time of a request i equal
to Max(0,li-ei). Li the latest time by which the
disk must begin servicing this request Ei the
earliest time at which the disk may begin
servicing a request
Different class-specific schedulers have
different insert strategies. ?Real-time
EDFSCAN ?Interactive Steal slack
time ?Throughput-intensive insert towards the
tail


soft real-time r3
interactive
Scheduled Queue i3 i2 i1
i3 r2 i2 i1 r1
throughput-intensive t1
Unused disk bandwidth for interactive class can
be assigned to other classes according to their
weight.
soft real-time r3 interactive
Scheduled Queue
t1 r3 i3 r2
i2 i1 r1


throughput-intensive t1
23
Cello Disk Scheduler Evaluation(1)
24
Cello Disk Scheduler Evaluation(2)
25
Open Discussions Section Two

• Overload of recomputing slack time after
  insertion of each request.
• Possible reductions?

26
Lazy Receiver Processing (LRP)
27
Conventional Linux Network Architecture for
Receiving Packets
28
Problem Analysis

• Eager receiver processing
• ?starving applications problem
• Lack of effective load shedding
• ?livestock problem
• Lack of traffic separation
• ?applications isolation problem
• Inappropriate resource accounting
• ?Account CPU time unfair

29
LRP Architecture
30
Features of LRP

• Lazy receiver processing
• Dropping packets as soon as possible
• Different NI channels for different sockets
• accurate resource accounting

31
Comparision of Two Architectures(1)
Demultiplexing Income Packets
32
Comparision of Two Architectures(2)
Packets Processing
33
Open Discussions Section
Three

• Accounting CPU time
• ?Virtual time
• How to make up the stolen time for a
  process?
• ?Lazy receiver processing

34
Related Works
35
QLinux
? Resource Reservation for different
classes of applications by multiple level
schedulers ? Reallocating dynamically
unused resources
36
ControlWare QLinux
A Problem with QLinux how to map QoS directly
to the appropriate weight? QLinux can only tell
that you can get some performance guarantee with
a certain weight. But if you want to get some
certain QoS, what is the appropriate weight
assigned to you?
A possible solution using some feedback
mechanism like ControlWare to adjust the weights
until finding the appropriate weight for a
certain QoS.
37
QuO QLinux
QuO solves how to transmit the QoS information
between a client and an object based on the QoS
parameters of underlying OS and network. So
QLinux can serve very well for QuO.
38
References 1 P. Goyal and X. Guo and H.M. Vin,
A Hierarchical CPU Scheduler for Multimedia
Operating Systems, Proceedings of 2nd Symposium
on Operating System Design and Implementation
(OSDI'96), Seattle, WA, pages 107-122, October
1996. 2 P. Goyal and H. M. Vin and H. Cheng,
Start-time Fair Queuing A Scheduling Algorithm
for Integrated Services Packet Switching
Networks, Proceedings of ACM SIGCOMM'96, pages
157-168, August 1996. 3 P Shenoy and H M. Vin,
Cello A Disk Scheduling Framework for Next
Generation Operating Systems, Proceedings of ACM
SIGMETRICS Conference, Madison, WI, pages 44-55,
June 1998. 4 P. Druschel and G. Banga, Lazy
Receiver Processing (LRP) A Network Subsystem
Architecture for Server Systems, Proceedings of
the 2nd Symposium on Operating System Design and
Implementation (OSDI'96), Seattle, WA, Pages
261-275, October 1996 5 Vijay Sundaram,
Abhishek Chandra, Pawan Goyal and Prashant
Shenoy, Application Performance in the QLinux
Multimedia Operating System , Proceedings of the
Eighth ACM Conference on Multimedia, Los Angeles,
CA, pages 127-136, November 2000. November
2000. 6 John Regehr and John A. Stankovic,
HLSA Framework for Composing Soft Real-Time
Schedulers, Proceedings of the 22nd IEEE
Real-Time Systems Symposium (RTSS 2001), pages
3-14, London, UK, December 3-6 2001. 7 John
Regehr and John A. Stankovic, Augmented CPU
Reservations Towards Predictable Execution on
General-Purpose Operating Systems, Proceedings of
the Seventh IEEE Real-Time Technology and
Applications Symposium (RTAS 2001), pages
141-148, Taipei, Taiwan, May 30-June 1 2001 8
John Regehr and John A. Stankovic, Operating
System Support for Multimedia The Programming
Model Matthers