Chapter 9 – Cloud Resource
Management and Scheduling
Contents
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
Policies and mechanisms for resource management.
Cloud resource utilization.
Resource management and dynamic application scaling.
Control theory and optimal resource management.
A two-level resource allocation architecture.
Feedback control based on dynamic thresholds.
Coordination of power and performance management.
A utility-based model for cloud-based Web services.
Scheduling algorithms for computer clouds.
Delay scheduling.
Data-aware scheduling.
Apache capacity scheduler.
Start-up fair queuing.
Borrowed virtual time.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
2
Policies and mechanisms for resource management
◼
◼
Resource management → critical function of any man-made system.
Such systems have finite resources that have to be shared among
different system components.
Policies and mechanisms for resource allocation.
 Policy → principles guiding decisions.
 Mechanisms → the means to implement policies.
◼
Resource management affects three basic criteria for system evaluation:
 Functionality → if the system functions according to specification.
 Performance → if the system design performance criteria are met.
 Cost → if the cost of building and maintaining the system meet specifications.
◼
Scheduling in a computing system → deciding how to allocate resources
of a system, such as CPU cycles, memory, secondary storage space, I/O
and network bandwidth, between users and tasks.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
3
Challenges for cloud resource management
◼
CRM - Cloud resource management:
 Requires complex policies and
decisions for multi-objective optimization.
 Affected by unpredictable interactions with the environment, e.g., system
failures, attacks.
 Cloud service providers are faced with large fluctuating loads which
challenge the claim of cloud elasticity.
◼
Effective CRM is extremely challenging;
The scale of the cloud infrastructure → makes it impossible to have accurate global
state information
 The interactions of the system with a large user population → makes it nearly
impossible to predict the type and the intensity of the system workload.

◼
The strategies for resource management for IaaS, PaaS, and SaaS are
different.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
4
CRM policies
1.
2.
3.
4.
5.
Admission control → prevent the system from accepting workload
in violation of high-level system objectives.
Capacity allocation → allocate resources for individual activations
of a service.
Load balancing → distribute the workload evenly among the
servers.
Energy optimization → minimization of energy consumption.
Quality of service (QoS) guarantees → ability to satisfy timing or
other conditions specified by a Service Level Agreement.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
5
CRM mechanisms
◼
◼
◼
◼
Control theory → uses the feedback to guarantee system stability
and predict transient behavior.
Machine learning → does not need a performance model of the
system.
Utility-based → require a performance model and a mechanism to
correlate user-level performance with cost.
Market-oriented/economic → do not require a model of the system,
e.g., combinatorial auctions for bundles of resources.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
6
Tradeoffs
The normalized performance and energy consumption, function of the
processor speed; the performance decreases at a lower rate than
does the energy when the clock rate decreases.
To reduce cost and save energy we may need to concentrate the load
on fewer servers rather than balance the load among them.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
7
Cloud resource utilization and energy efficiency
◼
Energy use for computing scales linearly with the number of computing
devices. Indeed, performance growth rate and improvements in electrical
efficiency almost cancel out.
 According to Moore's Law the number of transistors on a chip, thus, the
computing power of microprocessors doubles every 1.5 years.
 Electrical efficiency of computing devices doubles about every 1.5 years.
◼
◼
The energy consumption of cloud data centers is growing, has a
significant ecological impact, and affects the cost of cloud services.
Cloud services costs are affected by energy costs.
 Example - the costs for two AWS regions, US East and South America are:
upfront $2,604/year versus $5,632/year; hourly cots are $0.412 versus
$0.724.
 Higher energy and communication costs are responsible for the difference in
this example; the energy costs for the two regions differ by about 40% .
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
8
Cloud elasticity and overprovisioning
◼
◼
◼
Elasticity → additional resources are guaranteed to be allocated
when an application needs them and these resources will be
released when they are no longer needed. The user ends up paying
only for the resources it has actually used.
Over-provisioning → a cloud service provider has to invest in a
larger infrastructure than the typical cloud workload warrants. It
follows that the average cloud server utilization is low.
Elasticity is based on overprovisioning and on two assumptions:
 There is an effective admission control mechanism.
 The likelihood of all running applications dramatically increasing their
resource consumption at the same time is extremely low.
◼
◼
◼
Performance per Watt of power (PWP) → common measure of
energy efficiency.
Low server utilization affects negatively PWP and the ecological
impact of cloud computing.
Conclusion →overprovisioning is not economically sustainable.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
9
Percentage of
power usage
100
Typical operating
region
90
Power
80
70
Energy
efficiency
60
50
40
30
20
10
0
0
10
20
30
40
50
60
70
80
90
100
Percentage
of system
utilization
Even when power requirements scale linearly with the load, the energy efficiency of
a computing system is not a linear function of the load. Even when idle, a system
may use 50\% of the power corresponding to the full load. Typical operating region
for the servers at a data center is from about 10% to 50% of the load.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
10
Energy efficiency and energy-proportional systems
◼
◼
◼
◼
◼
An energy-proportional system consumes no power when idle, very
little power under a light load and more power as the load increases.
The system always operates at 100% efficiency.
The dynamic range is determined by the lower and the upper limit of
the device power consumption. A large dynamic range means that
the device is able to operate at a lower fraction of its peak power
when its load is low.
Different subsystems of a computing system behave differently in
terms of energy efficiency.
Processors used in cloud servers consume less than 1/3 of their
peak power at very-low load and have a dynamic range of more
than 70% of peak power. Processors used in mobile and/or
embedded systems are better.
Example: a 2.4 GHz Intel Q6600 processor with 4 GB of RAM
consumes 110 W when idle and 175 W when fully loaded.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
11
Energy saving
◼
◼
◼
◼
The alternative to resource management policy when servers are
always on, regardless of their load, is to develop energy-aware load
balancing and scaling policies.
Such policies combine dynamic power management with load
balancing and attempt to identify servers operating outside their optimal
energy regime and decide if and when they should be switched to a
sleep state or what other actions should be taken to optimize the energy
consumption.
Energy optimization cannot be considered in isolation, it has to be
coupled with admission control, capacity allocation, load balancing, and
quality of service. Existing mechanisms cannot support concurrent
optimization of all the policies.
Mechanisms based on a solid foundation such as control theory are too
complex and do not scale well. Those based on machine learning are
not fully developed, and the others require a model of a system with a
dynamic configuration operating in a fast-changing environment.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
12
Resource management and dynamic scaling
◼
Two application scaling strategies:
 Vertical scaling → keeps the number of VMs of an application constant,
but increases the amount of resources allocated to each one of them.
This can be done either by migrating the VMs to more powerful servers,
or by keeping the VMs on the same servers, but increasing their share
of the CPU time. The first alternative involves additional overhead; the
VM is stopped, a snapshot of it is taken, the file is transported to a
more powerful server, and, finally, the VM is restated at the new site.
 Horizontal scaling → common scaling strategy on a cloud; it is done by
increasing the number of VMs as the load increases and reducing this
number when the load decreases. Often, this leads to an increase of
communication bandwidth consumed by the application. Load balancing
among the running VMs is critical for this mode of operation.
◼
An application should de designed to support scaling.
 Workload partitioning of a modularly divisible application is static.
 The workload
of an arbitrarily divisible application can be partitioned
dynamically.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
13
Control theory and optimal CRM
◼
◼
◼
Control theory has been used to design adaptive resource
management for several classes of applications including power
management, task scheduling, QoS adaptation in web servers, and
load balancing.
Classical feedback control methods are used to regulate key operating
parameters of the system based on measurement of system output.
The main components of a cloud control system:
 Inputs → offered workload and policies
for admission control, capacity
allocation, load balancing, energy optimization, and the QoS guarantees.
 Control system components → sensors used to estimate relevant
measures of performance and controllers which implement policies.
 Outputs → the resource allocations to the individual applications.
◼
The feedback control assumes a linear time-invariant system model,
and a closed-loop controller. This controller is based on an open-loop
system transfer function which satisfies stability and sensitivity
constraints.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
14
Feedback and Stability
◼
Control granularity → the level of detail of the information used to
control the system.
 Fine control → very detailed information about the parameters
controlling the system state is used.
 Coarse control → the accuracy of these parameters is traded for the
efficiency of implementation.
◼
◼
The controllers use the feedback provided by sensors to stabilize
the system. Stability is related to the change of the output.
Sources of instability in any control system:
 The delay in getting the system reaction after a control action.
 The granularity of the control, the fact that a small change enacted by
the controllers leads to very large changes of the output.
 Oscillations, when the changes of the input are too large and the
control is too weak, such that the changes of the input propagate
directly to the output.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
15
The structure of a cloud controller
disturbance
r
external
traffic
Predictive
filter
forecast
Optimal
controller
 (k )
s
u* (k)
Queuing
dynamics
 (k )
state feedback q(k)
The controller uses the feedback regarding the current state and the estimation
of the future disturbance due to environment to compute the optimal inputs over
a finite horizon. r and s are the weighting factors of the performance index.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
16
A two-level resource allocation architecture
◼
The automatic resource management is based on two levels of
controllers, one for the service provider and one for the application.
 Inputs are: the offered workload and the policies
for admission control,
the capacity allocation, the load balancing, the energy optimization, and
the QoS guarantees in the cloud.
 System components are sensors used to estimate relevant measures of
performance and controllers implementing various policies.
 Output is the resource allocations to the individual applications.
◼
Is it beneficial to have two types of controllers:
 application controllers → determine if additional resources are needed.
 cloud controllers → arbitrate requests for resources and allocates the
physical resources.
◼
◼
◼
Choose fine versus coarse control.
Dynamic thresholds based on time averages better versus static ones.
Use a high and a low threshold versus a high threshold only.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
17
Two-level cloud controller
Application n
Application 1
Application 1
SLA 1
Application
controller
…. SLA n
….
Application n
Application
controller
VM
VM
Monitor
Decision
Cloud Controller
….
Actuator
VM
VM
Monitor
….
Decision
Actuator
Cloud Platform
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
18
Lessons from the two-level experiment
◼
◼
◼
◼
The actions of the control system should be carried out in a rhythm
that does not lead to instability.
Adjustments should only be carried out after the performance of the
system has stabilized.
If upper and a lower thresholds are set, then instability occurs when
they are too close to one another if the variations of the workload
are large enough and the time required to adapt does not allow the
system to stabilize.
The actions consist of allocation/deallocation of one or more virtual
machines. Sometimes allocation/dealocation of a single VM
required by one of the threshold may cause crossing of the other,
another source of instability.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
19
More on control theory application to CRM
◼
◼
◼
◼
◼
Regulate the key operating parameters of the system based on
measurement of the system output.
The system transfer function satisfies stability and sensitivity
constraints.
A threshold → the value of a parameter related to the state of a
system that triggers a change in the system behavior.
Thresholds → used to keep critical parameters of a system in a
predefined range.
Two types of policies:
threshold-based → upper and lower bounds on performance trigger
adaptation through resource reallocation; such policies are simple and
intuitive but require setting per-application thresholds.
2. sequential decision → based on Markovian decision models.
1.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
20
Feedback control based on dynamic thresholds
◼
Algorithm
 Compute the integral value of the high and the low threshold as
averages of the maximum and, respectively, the minimum of the
processor utilization over the process history.
 Request additional VMs when the average value of the CPU utilization
over the current time slice exceeds the high threshold.
 Release a VM when the average value of the CPU utilization over the
current time slice falls below the low threshold.
◼
Conclusions
 Dynamic thresholds perform better than the static ones.
 Two thresholds are better than one.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
21
Coordination of power and performance management
◼
◼
◼
◼
◼
◼
Use separate controllers/managers for the two objectives.
Identify a minimal set of parameters to be exchanged between the
two managers.
Use a joint utility function for power and performance.
Set up a power cap for individual systems based on the utilityoptimized power management policy.
Use a standard performance manager modified only to accept
input from the power manager regarding the frequency determined
according to the power management policy.
Use standard software systems.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
22
Communication between
autonomous managers
Performance manager
Control policy
Power manager
Control policy
Power
data
Performance
data
Blade
Blade
Workload
generator
Workload
distribution
Blade
Power
assignment
Power
Blade
Blade
Autonomous performance and power managers cooperate to
ensure prescribed performance and energy optimization; they
are fed with performance and power data and implement the
performance and power management policies
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
23
Utility-based model for cloud-based web services
◼
◼
◼
◼
◼
A service level agreement (SLA) → specifies the rewards as well
as penalties associated with specific performance metrics.
The SLA for cloud-based web services uses the average response
time to reflect the Quality of Service.
We assume a cloud providing K different classes of service, each
class k involving Nk applications.
The system is modeled as a network of queues with multi-queues
for each server.
A delay center models the think time of the user after the
completion of service at one server and the start of processing at
the next server.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
24
Utility function U(R)
P
e
n
a
t
l
t
y
R
e
w
a
r
d
0
R0
R1
R2
R - response time
The utility function U(R) is a series of step functions with jumps
corresponding to the response time, R=R0 | R1 | R2, when the
reward and the penalty levels change according to the SLA. The
dotted line shows a quadratic approximation of the utility function.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
25
Si2
Si3
Si1
Si6
Si4
vk
vkmax
Si5
rkmax
rk
(b)
(a)
(a) The utility function: vk the revenue (or the penalty)
function of the response time rk for a request of class k.
(b) A network of multiqueues.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
26
The model requires a large number of parameters
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
27
Cloud scheduling algorithms
◼
Scheduling → responsible for resource sharing at several levels:
 A server can be shared among several virtual machines.
 A virtual machine could support several applications.
 An application may consist of multiple threads.
◼
◼
A scheduling algorithm should be efficient, fair, and starvation-free.
The objectives of a scheduler:
 Batch system → maximize throughput
and minimize turnaround time.
 Real-time system → meet the deadlines and be predictable.
◼
◼
Best-effort: batch applications and analytics.
Common algorithms for best effort applications:
 Round-robin.
 First-Come-First-Serve (FCFS).
 Shortest-Job-First (SJF).
 Priority algorithms.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
28
More on cloud scheduling algorithms
◼
Multimedia applications (e.g., audio and video streaming)
 Have soft real-time constraints.
 Require statistically guaranteed maximum delay and throughput.
◼
◼
Real-time applications have hard real-time constraints.
Scheduling algorithms for real-time applications:
 Earliest Deadline First (EDF).
 Rate Monotonic Algorithms (RMA).
◼
Algorithms for integrated scheduling of several classes of
applications:
 Resource Allocation/Dispatching (RAD) .
 Rate-Based Earliest Deadline (RBED).
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
29
Quantity
Hard real-time
strict
Hard-requirements
Soft-requirements
Best-effort
Timing
loose
loose
strict
Best-effort policies → do not impose requirements regarding either the amount of
resources allocated to an application, or the timing when an application is scheduled.
Soft-requirements policies → require statistically guaranteed amounts and
timing constraints
Hard-requirements policies → demand strict timing and precise amounts of resources.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
30
Delay scheduling
◼
◼
◼
How to simultaneously ensure fairness and maximize resource
utilization without compromising locality and throughput for Big Data
applications running on large computer clusters?
This was one of the questions faced early on in the cloud computing
era by Facebook, Yahoo, and other large IT service providers.
Example - each Hadoop job consists of multiple Map and Reduce
tasks and the question is how to allocate resources to the tasks of
newly submitted jobs. Hadoop approach:
 The job tracker of the Hadoop master manages a number of slave
servers running under the control of task trackers with slots for Map
and Reduce tasks.
 A FIFO scheduler with five priority levels assigns slots to tasks based on
their priority.
 The fewer the number of tasks of a job already running on slots of all
servers, the higher is the priority of the remaining tasks.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
31
Challenges of priority based mechanisms
◼
◼
Priority-based allocation does not consider data locality, namely the
need to place tasks close to their input data.
Locality affects the throughput,
 Server locality, i.e., getting data from the local server, is significantly
better in terms of time and overhead than rack locality, i.e. getting input
data from a different server in the same rack.
 The network bandwidth in a large cluster is considerably lower than the
disk bandwidth; also the latency for local data access is much lower
than the latency of a remote disk access.
◼
◼
In steady-state, priority scheduling leads to the tendency to assign
the same slot repeatedly to the next task(s) of the same job. As one
of the job's tasks completes execution its priority decreases and the
available slot is allocated to the next task of the same job.
Priority scheduling favors the occurrence of sticky slots!!
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
32
Task locality and average job locality
◼
◼
◼
A task assignment satisfies the locality requirement if the input task
data are stored on the server hosting the slot allocated to the task.
How should a fair scheduler operate on a shared cluster? What is the
number n of slots of a shared cluster the scheduler should allocate to
jobs assuming that tasks of all jobs take an average of T seconds to
complete?
A sensible answer is that the scheduler should provide enough slots
such that the response time on the shared cluster should be the same
as the completion time of the job on a fictitious private cluster with n
available slots for the n tasks of the job as soon as job arrives.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
33
Delay scheduling a counterintuitive policy
◼
◼
◼
◼
This policy delays scheduling the tasks of a new job for a relatively
short time to address the conflict between fairness and locality.
The new policy skips a task of the job at the head of the priority
queue if the input data are not available on the server where the slot
is located and repeats this process up to D times as specified by the
delay scheduling algorithm showed in the next box.
Delay scheduling performs well when most tasks are short relative
to job duration, and when a running task can read a given data block
from multiple locations.
Results for workloads at Yahoo and Facebook show an almost
doubling of the throughput under the new policy, while ensuring
fairness.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
34
Hadoop fair scheduler (HFS)
◼
HFS design goals:
1.
2.
3.
Fair sharing at the level of users rather than jobs. This requires a twolevel scheduling, the first level allocates task slots to pools of jobs
using a fair sharing policy; at the second level each pool allocates its
slots to jobs in the pool.
User controlled scheduling; the second level policy can be either FIFO
or fair sharing of the slots in the pool.
Predictable turnaround time. Each pool has a guaranteed minimum
share of slots. To accomplish this goal HFS defines a minimum share
timeout and a fair share timeout and when the corresponding timeout
occurs it kills buggy jobs or tasks taking a very long time. Instead of
using a minimum skip count it use a wait time to determine how long a
job waits to allocate a slot to its next ready-to-run task.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
35
HFS operation
◼
◼
◼
◼
◼
◼
HFS creates a sorted list of jobs ordered according to its scheduling
policy.
Scans down this list to identify the job allowed to schedule a task next,
and within each pool applies the pool's internal scheduling policy.
Pools missing their minimum share are placed at the head of the sorted
list and the other pools are sorted to achieve a weighted fair sharing.
A job starts at locality level 0 and can only launch node-local tasks.
After at least W1 seconds the job advances at level 1 and may launch
rack-local tasks, then after a further W2 seconds, it goes to level 2 and
may launch off-rack tasks.
If a job launches a local task with a locality higher than the level it is on,
it goes back down to a previous level.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
36
9.11 Data-aware scheduling
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
37
Apache capacity scheduler
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
38
Fair queuing and startup fair queuing
R(t)
R(t)
Fai(tai)=Sai(tai)+Pai
Fai(tai)=Sai(tai)+Pai
Sai(tai)=Rai(tai)
Sai(tai)=Fai-1(tai-1)
Fai-1(tai-1)
Rai(tai)
tai
tai-1
tai-1
(a)
tai
(b)
Fair queuing - schedule multiple flows through a switch. The transmission of
packet i of a flow can only start after the packet is available and
transmission of the previous packet has finished.
(a) The new packet arrives after the previous has finished.
(b) The new packet arrives before the previous one was finished.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
39
Start-time fair queuing (SFQ)
◼
◼
Organize the consumers of the CPU bandwidth in a tree structure.
The root node is the processor and the leaves of this tree are the
threads of each application.
 When a virtual machine is not active, its bandwidth is reallocated to the
other VMs active at the time.
 When one of the applications of a virtual machine is not active, its
allocation is transferred to the other applications running on the same VM.
 If one of the threads of an application is not runnable then its allocation is
transferred to the other threads of the applications.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
40
Server
VM2
(3)
VM1
(1)
A2
(1)
A1
(3)
t1,1
(1)
t1,2
(1)
t1,3
(1)
t2
(1)
A3
(1)
vs1
(1)
vs2
(1)
vs3
(1)
The SFQ tree for scheduling when two virtual machines VM1 and VM2 run
on a powerful server.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
41
Borrowed virtual time (BVT)
◼
◼
Objective - support low-latency dispatching of real-time applications,
and weighted sharing of CPU among several classes of applications.
A thread i has
 an effective virtual time, Ei.
 an actual virtual time, Ai.
 a virtual time warp, Wi.
◼
◼
◼
The scheduler thread maintains its own scheduler virtual time (SVT)
defined as the minimum actual virtual time of any thread.
The threads are dispatched in the order of their effective virtual time,
policy called the Earliest Virtual Time (EVT).
Context switches are triggered by events such as:
 the running thread is blocked waiting for an event to occur.
 the time quantum expires.
 an interrupt occurs.
 when a thread becomes runnable after sleeping.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
42
Thread a
0
12
24
36
t
12
24
36
thread b is suspended
48
thread b is reactivated
Thread b
0
3
3
6
6
9
9
12
12
15
Virtual time
36
24
12
Real time
6
0
3
15
18
21
24
36
48
60
Top → the virtual startup time and the virtual finish time and function of the real time t
for each activation of threads a and b.
Bottom → the virtual time of the scheduler v(t) function of the real time
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
43
Effective virtual time
450
390
360
300
270
210
180
120
90
30
Real time (mcu)
2
5
11
9
14
20
18
23
29
27
32
38
36
41
45
The effective virtual time and the real time of the threads a
(solid line) and b (dotted line) with weights wa = 2 wb when
the actual virtual time is incremented in steps of 90 mcu.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
44
Effective virtual time
450
390
360
300
270
210
180
120
90
30
Real time (mcu)
2
5
12
9
14
21
18
23
30
27
32
36
39
41
45
-60
The effective virtual time and the real time of the threads a (solid line), b (dotted line),
and the c with real-time constraints (thick solid line). Thread c wakes up periodically at
times t=9, 18, 27, 36,…, is active for 3 units of time and has a time warp of 60 mcu.
Dan C. Marinescu
Cloud Computing Third Edition - Chapter 9
45