Service Provisioning based on Advanced Reservation and Deadline Sensitive Leases

In IaaS Cloud

N.Tajunnisha, S. Jaya Nirmala, S. Mary Saira Bhanu,

CSE Department, National Institute of Technology,
Tiruchirappalli, Tamilnadu, India.
e-mail: [email protected], [email protected], [email protected]

Abstract— The need for computational resources are resources and storage as standardized services for users over
increasing day by day and has become a fundamental the network. It is important to note that SaaS is built on
requirement in industry. IaaS cloud provides computing PaaS, and the latter on IaaS [2]. .
resources and storage as standardized services for users The need for computational resources are increasing day
over the network. Computational resource provisioning by day and has become a fundamental requirement in
has three dimensions: hardware resource, software industry. Lease based approaches and job based approaches
stack available on those resources, and the time during are two different approaches used for provisioning
which those resources are available to users. Lease based computational resources [9]. Resource provisioning has
approaches and job based approaches are two different been mostly tied to running jobs and many job schedulers
approaches used for provisioning computational have been developed such as :Maui [12], Moab [13],
resources. Virtual machines, a lease based approach can LoadLeveler[14] and so on. Job-based systems provision
be used to provision resources across these dimensions resources where the user does not have direct access to the
and therefore considered as an appealing vehicle for resources but is only allowed to submit jobs. These systems
computational resource provisioning. Different possible can support both best-effort and advance reservation
types of service request for IaaS cloud are advanced provisioning of resources, but have limited support for
reservation, immediate, deadline sensitive and best- custom software environments. Although many modern
effort. A service provisioning mechanism which is a lease schedulers support check-pointing based preemption, the
based approach is proposed that will schedule the application must itself be check-pointable. An application
different types of service requests in a IaaS cloud. The can be made check-pointable by explicitly adding that
proposed algorithm is implemented in Java that also functionality to an application or by using OS-level check-
enacts the schedule by issuing OpenNebula commands to pointing, where the operating system checkpoints a process.
launch, terminate, suspend, resume the virtual machines This burdens the user with having to modify their
and manage the computational resources. Experiments applications to make them check-pointable.
are conducted and results show that the proposed A lease is a form of contract where the provider agrees to
algorithm can schedule more service requests than provide a set of resources to the consumer under a set of
existing approaches and the system utilization is terms specifying the time duration, amount of resources,
improved because of reduced resource fragmentation. access to the resources, etc. In the context of leasing
Keywords- Cloud computing, computational resources, computational resources, these terms involve details about
virtual machines, IaaS cloud, OpenNebula, resource the hardware resources, such as CPU, memory, network
fragmentation. bandwidth, software environment required and an
availability period during which the requested resources
I. INTRODUCTION must be made available. Computational resource
Cloud computing provides a large pool of computing provisioning has three dimensions: hardware resource,
resources with dynamically scalable infrastructure for software stack available on those resources, and the time
application and data storage. With cloud computing, the cost during which those resources are available to users. A
of application hosting, information storage and retrieval is complete resource provisioning model should allow
reduced significantly. The National Institute of Standards resource consumers to specify their requirements in these
and Technology [1] defines cloud computing as “a model three dimensions, and the resource provider to satisfy the
for enabling convenient, on-demand network access to a specified requirements. Virtual machines, a lease based
shared pool of configurable computing resources (eg., approach can be used to provision resources across these
networks, servers, storage, applications, and services) that dimensions and therefore considered as an appealing vehicle
can be rapidly provisioned and released with minimal for computational resource provisioning [9].
management effort or service provider interaction”. Cloud Virtual machines can be mapped to all or part of physical
computing service provisioning models based on capability hardware resources, allowing users for fine-grained
are IaaS, PaaS and SaaS. IaaS provides computing specification of their hardware resource requirement. Virtual
machines also have the additional property of making these need to be provisioned immediately. For example, a
mappings strictly enforceable than a job based approach. telecommunications company having an existing
Virtual machine images can be customized with required infrastructure may need to supplement that infrastructure
software environment, allowing users to specify their with additional resources during periods of unforeseen
software requirement while job based systems have limited increased web traffic and application load, meaning those
or no support for providing custom software environments additional resources have to made available from the cloud
for each job request. VMs can be suspended, migrated, and provider with very little advance notice. In this case,
resumed with out modifying the applications running inside Performance can be measured in terms of number of
the VM. Although there are some other mechanisms in job accepted requests.
based resource provisioning models to suspend, resume or
B. Resource Allocation In IaaS Cloud
migrate computation (for example check-pointing and
preempting schedulers) VMs provide a more consistent Comparison of various cloud toolkits based on resource
solution because they do not require applications, or even allocation policies is given in Table 1. OpenNebula [3],
the OS running inside the VM, to be aware of these VM Nimbus [5] and Eucalyptus [6] are cloud toolkits that can be
operations. used to setup a cloud on local infrastructure. OpenNebula
This paper proposes a strategy for IaaS cloud for comes with a simple match making scheduler (mm_sched)
efficient service provisioning for different possible types of which is in charge of the assignment between pending
service requests. Rest of the paper is structured as follows: Virtual Machines and known physical hosts. The goal of
Section 2 discusses the background and related work. this policy is to prioritize the hosts more suitable for
Section 3 describes about the system model and the proposed deploying the upcoming VM and accept resource requests
scheduling methodology for service provisioning in IaaS on best-effort basis. Haizea is a resource lease manager [7]
cloud. Section 4 discusses about the experimental results and that can be used as a drop-in replacement for OpenNebula’s
Section 5 gives the conclusion and future work. default scheduling daemon(mm_sched). Amazon EC2 [4] is
a public cloud which provides computing resources on pay-
per-use model to general public.
II. BACKGROUND
IaaS cloud allocates resources to competing requests
In cloud, service provisioning is the automation of all based on predefined resource allocation policies. In the
the steps required to service user requests and manage the present scenario, most of the Infrastructure as a Service
physical resources. (IaaS) clouds use simple resource allocation policies like IM
and BE. For example, Eucalyptus supports IM resource
A. Types of Service Requests allocation policy while OpenNebula supports BE resource
The different types of possible service requests in a IaaS allocation policy. IM policy allocates the resources if it is
cloud [8] are Best-effort(BE), Deadline Sensitive(DS), available at the instant of time when the request is
Advanced Reservation(AR) and Immediate(IM). submitted, otherwise the request is rejected. BE resource
The resource consumer sometimes may need the allocation policy also allocates the requested resources if
resources for specified duration without strict time slots, available otherwise the request is placed in a FIFO queue.
which corresponds to BE service request. Performance of Cloud providers cannot satisfy the various types of service
BE requests can be measured in terms of response time. In requests by considering only BE or IM resource allocation
some cases, the resource consumer needs the resources for policies due to availability of finite resources at a time.
specified duration with additional deadline constraint which Table 1 Comparison of various cloud toolkits based on
is called as DS service request. For example, a scientist may resource allocation policies
require a large number of computers to run a experiment for
just a few hours, but might not need those computers at any S.No Cloud Toolkits Allocation Policies
specific time slots (as long as they are made available within 1 Eucalyptus Immediate
a reasonable amount of time). Performance of DS requests 2 Nimbus Immediate
can be measured in terms of both response time and number 3 Amazon EC2 Best-effort
of accepted requests. 4 OpenNebula Best-effort
The resource consumer needs the specified resources in 5 OpenNebula Immediate, best-effort, advance
fixed time slots in certain cases which is called as AR type with Haizea reservation
service request. For example, a college lab instructor may
need a cluster of computers available to students during the The user request computational resources in the form of
course's lab session which is at very specific times during a lease in Haizea. It will then reserve the resources for this
the week and with a specific software configuration. lease. Whenever start time of a reservation comes, resources
Performance of this type of service request can be measured are allocated to users in the form of VMs, which are
in terms of number of accepted requests. IM type is a deployed on physical machines. It assumes that BE leases
special form of AR request where the specified resources are preemptable and they neither have deadline nor any time
constraints. IM and AR leases are non-preemptable and memory state file is created on the node where the VM was
have time constraints, such as start time and end time. It will running. When a VM is resumed, memory state file is read
preempt BE leases whenever the resources are required for into memory. When a suspended VM is migrated to a
AR or IM leases. There is no guarantee that a submitted BE different node, its VM image and memory state file are
lease will get resources for completion within a certain time transferred to that node.
limit. The following conditions are assumed for the lease
Deadlines can be associated with BE leases for those requests based on the assumptions made in [11] and [8].
consumers who want to get resources for their BE leases • All the service requests are aperiodic
within a certain time limit. These kinds of leases can be • All requests are independent and the scheduling algorithm
considered as DS leases. Amit Nathani et al [8] proposed a assumes that there is no communication among the service
scheduling algorithm that supports four kinds of resource requests
allocation policies: IM, BE, AR and DS. On the arrival of a • DS and BE type requests are preemptible
lease request, it tries to schedule it in a single slot or • The amount of computational resources specified in the
multiple slots. If it is not possible then it tries to schedule request is the maximum limit that the request can use.
the request by modifying the existing schedule by using
swapping and backfilling (SAB) techniques. In swapping,
two consecutive leases are swapped if and only if, the first
lease has requested fewer resources than second lease and
after swapping their deadlines are not missed. Backfilling
procedure is applied to schedule BE and DS leases by
rescheduling them into multiple idle slots. SAB techniques
are applied to push idle resources towards requested time
slot of a new lease.
III. PROPOSED METHODOLOGY

A. System Description
Consumer has to submit lease request specified with the
following parameters.
• Amount of CPU
• Amount of memory in MB
• Start time Figure 1 Control Flow Diagram
• Duration
• Deadline
• Software to be deployed in allotted resources B. Proposed Scheduling Algorithm
A unique ‘ lease id’ will be assigned by the scheduler A Lease request is represented with the leasetype
for each submitted lease. 'start time' is the time when the datastructure in the following algorithm. When a new
submitted request is ready to accept the service. 'start time' lease request arrives, SCHEDULE(leasetype lease)
need to be specified for a BE or AR or IM request while for method is called.
a DS request, both 'start time' and ‘deadline’ need to be Lease requests are maintained in the sorted in the
specified. 'Duration' need to be specified for all the types of order of IM, AR, DS and BE request types. If two lease
service request. 'Arrival time' is the time at which the lease requests are of same request type, it is sorted by
request is submitted to the system. Schedule is prepared increasing order of deadline. If two lease requests are of
based on the available resources and the submitted lease same request type and deadline, it is sorted by the
requests. While enacting the schedule, resources are decreasing order of amount of computational resources
provisioned by launching virtual machines and are released needed by them. The scheduling policy followed here is
back to the system by terminating virtual machines after priority based where the parameters considered are
completion time. At any time, when a new lease request is request type, deadline and the amount of required
submitted, the system tries to accommodate it in the future computational resources. When a new lease request
schedule by rescheduling. If the system is able to find room arrives, it is inserted in the appropriate place in the list
for the new lease request by rescheduling, then it is accepted such that the above mentioned ordering is preserved.
else rejected. The software specified in the software field of 'Active lease' is the term used to describe the leases
the request is considered as the VM image. These VM whose deadline is beyond the current system time. The
images are stored at a central repository. If a lease is set of leases to be re-scheduled(LR) are in the above
preempted in schedule, it is suspended by suspending the mentioned sorted order when the method
corresponding virtual machine. When a VM is suspended, a RESCHEDULE(LR, new_schedule) is called. New
 schedule is created by scheduling them one after the
Algorithm: RESCHEDULE(ListOfLeases LR,Schedule
other by calling ADDLEASE(new_schedule ,lease).
new_schedule)
Algorithm: SCHEDULE(leasetype lease) 1 For each lease in LR
1 Insert the new lease request into the list of active sorted 1.1 ADDLEASE(new_schedule ,lease)
lease requests
2 If lease.type is AR
2.1 If a single idle slot is available that fits with
lease.CPU, lease.memory, lease.duration at the time Algorithm: ADDLEASE(Schedule new_schedule ,
point of lease.starttime in the schedule leasetype lease)
2.1.1 then book the slot by updating the schedule and 1 If lease.type is AR
accept the lease 1.1 If a single idle slot is available that fits with
3 else if lease.type is DS lease.CPU, lease.memory, lease.duration at the time
3.1 Find an idle slot that fits with lease.CPU, point of lease.starttime in new schedule
lease.memory in the schedule from the time point of 1.1.1 then book the slot by updating the new schedule
lease.starttime within lease.deadline and satifying and mark the lease as scheduled
the condition 1.2 else mark the lease as not scheduled
slot.duration > 2*(suspend_delay + resume_delay) 2 else if lease.type is DS
3.2 Search for multiple slots satisfying 3.1 until 2.1 Find an idle slot that fits with lease.CPU,
Sum(slot.duration) = lease.duration lease.memory in new schedule from the time point of
3.3 If adequate idle slots are available within lease.starttime within lease.deadline and satifying the
lease.deadline, book the slots and accept the lease condition
4 else if lease.type is BE slot.duration > 2*(suspend_delay + resume_delay)
4.1 Find idle slots that fits with lease.CPU, lease.memory 2.2 Search for multiple idle slots satisfying 2.1 until
in schedule from the time point of lease.starttime and Sum(slot.duration) = lease.duration
satifying the condition slot.duration > 2.3 If adequate idle slots are available within
2*(suspend_delay + resume_delay) lease.deadline
4.2 Search for multiple slots satisfying 4.1 until 2.3.1 then book those slots by updating the new
Sum(slot.duration) = lease.duration schedule and mark the lease as scheduled
4.3 If adequate idle slots are available, book the slots and 2.4 else mark the lease as not scheduled
accept the lease 3 else if lease.type is BE
5 Freeze the schedule starting from current system time to 3.1 Find an idle slot that fits with lease.CPU,
future time point of (system time + processing delay). lease.memory in new schedule from the time point of
While rescheduling, freezed part of the schedule is not lease.starttime and satifying the condition
disturbed. slot.duration > 2*(suspend_delay + resume_delay)
6 else if lease.type is IM 3.2 Search for multiple idle slots satisfying 3.1 until
6.1 Create a new schedule with the list of active sorted Sum(slot.duration) = lease.duration
lease requests 3.3 If adequate idle slots are available
6.2 If any of the already accepted leases seem to miss 3.3.1 then book those slots by updating the new
their deadline in new schedule, schedule and mark the lease as scheduled
then reject the new request. 3.4 else mark the lease as not scheduled
6.3 else update the current schedule by new schedule and 4 else if lease.type is IM
accept the lease request 4.1 If single idle slot that fits with lease.CPU,
7 If the new lease is not accepted yet, lease.memory, lease.duration and at the time
7.1 Set of leases to be rescheduled (LR) = set of leases point of lease.starttime in new schedule is available
starting from the new lease in the set of active sorted 4.1.1 then book the slot by updating the new schedule
lease requests and mark the lease as scheduled
7.2 Create a new schedule by calling 4.2 else mark the lease as not scheduled
RESCHEDULE(LR, new_schedule).
7.3 If none of the already accepted leases in LR misses
Lease requests are sorted based on the parameters- service
its deadline in new schedule,
request type, deadline and amount of computing resources.
7.3.1 then update the current schedule by new
IM requests have higher priority while BE requests have
schedule and accept the lease
least priority. BE lease can be preempted by AR lease.
7.4 Else reject the lease
Lease requests with upcoming nearby deadlines are
scheduled first. Leases requesting more resources are
scheduled before the leases requesting less resources so that
idle slots can be pushed into future time slots. Whenever a Table 3 Leases accepted, rejected and system utilization
new lease request arrives, either the slot for the new lease is for 22 leases
added to the existing schedule or rescheduling occurs Sets for 22 Algorithm Accepted Rejected Utilization
depending on the service request type and the deadline of leases
the newly arrived lease request.
Suspend delay, Resume delay and Creation delay are the Haizea 13 9 67.96
estimated maximum delay for suspending, resuming and Set 1
SAB 18 4 84.86
creating a virtual machine respectively. Processing delay
specifies the estimated time for rescheduling when a new Proposed 19 3 89.08
lease request arrives. This parameter is needed to freeze the Haizea 14 8 58.03
schedule from current system time up to the point of current Set 2
system time plus processing delay. Delay parameters are SAB 17 5 65.53
measured while enacting the schedule and the corresponding Proposed 18 4 71.12
average values are updated in the configuration file.
Haizea 16 6 75
IV. EXPERIMENTS AND RESULTS Set 3
SAB 18 4 77.77
Four systems are connected in a private LAN and
configured with OpenNebula 3.8. The cloud environment is Proposed 18 4 81.07
setup using OpenNebula. Each system has a 4GB RAM and Haizea 15 7 79.43
Intel core i3 processor at 3GHz configuration. Ubuntu Set 4
SAB 18 4 88.30
12.04(Precise Pangolin) is installed in all the systems and
the virtual machine manager is KVM. The solution is Proposed 20 2 80.09
implemented in Java embedded with OpenNebula
Haizea 14.5 7.5 70.10
commands. Experiment are tested with more number of
Average
systems also that gives similar kind of results as given here. SAB 17.75 4.25 79.11
Proposed 18.75 3.25 80.34
Table 2 Leases accepted, rejected and system utilization
for 13 leases
Table 4 Leases accepted, rejected and system utilization
Sets for 13 Algorithm Accepted Rejected Utilization
for 33 leases
leases
Sets for 33 Algorithm Accepted Rejected Utilization
Haizea 9 4 73.27 leases
Set 1
SAB 12 1 84.28 Haizea 24 9 72.70
Proposed 12 1 90.74 Set 1
SAB 29 4 82.4
Haizea 10 3 90.17 Proposed 32 1 89.74
Set 2
SAB 13 0 94.64 Haizea 20 13 65.48
Proposed 13 0 93.51 Set 2
SAB 25 8 69.46
Haizea 11 2 62.50 Proposed 29 4 85.07
Set 3
SAB 11 2 62.50 Haizea 28 5 68.81
Proposed 11 2 74.07 Set 3
SAB 32 1 75
Haizea 8 5 79.46 Proposed 32 1 73.9
Set 4
SAB 10 3 88.39 Haizea 24 9 61.16
Proposed 10 3 78.5 Set 4
SAB 32 1 66.73
Haizea 9.5 3.5 76.35 Proposed 33 0 72.77
Average
SAB 11.5 1.5 82.45 Haizea 24 9 67.03
Proposed 11.5 1.5 84.21 Average
SAB 29.5 3.5 73.39
Proposed 31.5 1.5 80.37
Table 5 Leases accepted, rejected and system utilization
for 43 leases
Sets for 43 Algorithm Accepted Rejected Utilization
leases
Haizea 28 15 64.8
Set 1
SAB 38 5 81.6
Proposed 41 2 94.12
Haizea 33 10 60.33
Set 2
SAB 41 2 67.75
Proposed 42 1 73.51
Haizea 20 23 61.67
Set 3
SAB 40 3 68.26
Proposed 40 3 71.11
Figure 2 Graph for Accepted Leases Vs Submitted Leases
Haizea 26 17 69.87
Set 4
SAB 43 0 82.53
Proposed 43 0 83.23
Haizea 26.75 16.25 64.16
Average
SAB 40.5 2.75 75.03
Proposed 41.5 1.5 80.49

The proposed algorithm is implemented in two modules

– Scheduler module and Enactment module. The Scheduler
module implements the proposed algorithm and generates
the schedule based on the available physical resources. The
generated schedule is given as input to the Enactment
module. The Enactment module follows the schedule by
issuing appropriate OpenNebula commands for creating,
suspending, resuming and terminating the virtual machines
at the specified point of time. Figure 3 Graph for Rejected Leases Vs Submitted Leases
In order to evaluate the performance of the proposed
algorithm to schedule service requests, it is compared with
the existing algorithm of Haizea[7] and the algorithm[8]
that uses SAB technique. The experiment is performed by
considering four sets of 13, 22, 33 and 43 leases
respectively. Each of the input parameters in all four sets
follows random distribution. They are taken from the input
data sets used to conduct experiments in [8]. All the four
sets are given as input and the parameters number of
accepted leases, number of rejected leases and system
utilization are measured for each input set. This is repeated
for all three methods -Haizea, SAB algorithm and the
proposed algorithm, whose results are given in Tables 2,3,4
and 5. The corresponding average values are calculated for
each set for each method and plotted as graphs.
From the graph in Figure 2, it can be observed that the
proposed algorithm provides higher request acceptance rate
relative to Haizea and a slight increase relative to SAB
Algorithm.
Figure 4 Graph for System Utilization Vs Submitted Leases
 From the graph in Figure 3, it can be observed that the the clouds: a Berkeley view of cloud computing, Technical
Report No. UCB/EECS-2009-28, University of California at
proposed algorithm rejects lesser number of leases when
Berkley, USA, pp.1-23, 2009.
compared with Haizea and SAB Algorithm.
System utilization is the ratio of used resources to total [3] OpenNebula, http://www.opennebula.org/
resources. From the graph in Figure 4, it can be observed
that in the proposed algorithm provides better and consistent [4] Amazon EC2, http://aws.amazon.com/ec2/
system utilization when compared with Haizea and SAB
[5] Nimbus, http://www.nimbusproject.org/
algorithm.
[6] D. Nurmi, R. Wolski, C. Grzegorczyk, G. Obertelli, S. Soman,
V. CONCLUSION AND FUTURE WORK L. Youseff, D.Zagorodnov, The eucalyptus open-source cloud-
Thus, different types of service requests are scheduled computing system, in: Cloud Computing and Applications 2008
(CCA08), 2008.and applications conference, pages 102-106,
and provisioned by using the proposed scheduler and 2012.
enactment is done with the OpenNebula Cloud. Delays due
to Virtual Machine suspension and migration are considered [7] B.Sotomayor, R.Santiago Montero, I.Martín Llorente, I.Foster,
while scheduling to ensure that the time overhead in Virtual Infrastructure Management in Private and Hybrid
Clouds, IEEE Internet computing, vol. 13, no. 5, pp. 14-22,
switching the resources among different leases do not Sep./Oct. 2009.
dominate the duration of scheduled slots. Processing delay
for rescheduling is computed and the current schedule is [8] Amit Nathani, Sanjay Chaudhary, Gaurav Somani, Policy based
modified accordingly. Allocation of multiple slots to BE resource allocation in IaaS cloud, Future Generation Computer
and DS leases had resulted in improved request acceptance Systems 28 (2012) 94-103.
rate, reduced resource fragmentation and improved system
utilization. [9] Borja sotomayor, provisioning computational resources using
virtual machines and leases,
Future work is to consider the Service Level http://people.cs.uchicago.edu/~borja/dissertation/, August 2010.
Agreements (SLA) between the IaaS cloud provider and
consumers while provisioning the resources. User defined [10] Dhariya vora, Sanjay Chaudhary, Minal Bhise, Vikas Kumar:
Quality of Service (QoS) requirements such as budget Allocation of Slotted Deadline Sensitive Leases in Infrastructure
Cloud, Distributed Computing and Internet Technology,
(maximum cost payable for using the services) and Volume 7154, 2012, pp 242-252, 2012.
reliability (minimum reliability factor required for that
request) can be taken into account while scheduling service [11] B. Sotomayor, K. Keahey, I. Foster, Combining Batch
requests. Execution and Leasing Using Virtual Machines, in: HPDC,
Manchester, 2008, pp. 1-9.

[12] http://www.clusterresources.com/pages/products/maui-cluster-
REFERENCES
scheduler.php/
[1] P. Mell and T. Grance, The NIST definition of cloud
computing, National Institute of Standards and Technology, [13] http://www.clusterresources.com/pages/products/moab-cluster-
2009. suite/workload-manager.php

[2] Michael Armbrust, Armando Fox, Rean Griffith, Anthony D. [14] http://www.ibm.com/systems/clusters/software/loadleveler
Joseph, Randy H. Katz, Andrew Konwinski, Gunho Lee, David .html
A. Patterson, Ariel Rabkin, Ion Stoica, Matei Zaharia, Above