Enterprise application in the cloud – virtualized deployment

Enterprise Applications in the Cloud:

Virtualized Deployment

Leonid Grinshpan, Oracle Corporation (www.oracle.com)

Subject

Enterprise applications (EA) can be deployed in the Cloud in two ways:

1. Non-virtualized setup hosts on the same physical servers different EAs without
logical borders between them (no partitions, virtual machines or similar
technologies in place).

2. Virtualized arrangement separates EAs from each other by employing the
above-mentioned techniques.

In the article [Leonid Grinshpan. Enterprise Applications in the Cloud: Non-virtualized
Deployment; Measure IT, issue 10, 2011] we analyzed interference among EAs when they
are set up on physical servers without logical boundaries. We also outlined how to
collect data on resource utilization by hosted EAs. The current article studies the impact
of hardware virtualization on EA performance. We are using queuing models of EAs as
scientific instruments for our research; methodological foundation for EA performance
analysis based on queuing models can be found in the author’s book [Leonid Grinshpan.
Solving Enterprise Applications Performance Puzzles: Queuing Models to the Rescue, Willey-
IEEE Press, 2012, http://www.amazon.com/Solving-Enterprise-Applications-Performance-
Puzzles/dp/1118061578/ref=ntt_at_ep_dpt_1].

We begin by examining EA performance on a non-virtual platform that includes Web,
Application, and Database servers. In the next step, we substitute the Database server
with three virtual machines (VM) and compare EA performance on virtualized versus
non-virtualized platforms. Next, we evaluate how many CPUs should be assigned to

VMs to ensure acceptable transaction response times. In the last step, we discuss why
virtualization does not prevent applications from impacting each other’s performance.

Model 1 Application Performance on Physical Platform

For analysis we are using the queuing model from the article [Leonid Grinshpan.
Enterprise Applications in the Cloud. Non-virtualized Deployment; Measure IT, issue 10, 2011]
with slightly different workload and transaction profiles. Model 1 (Figure 1) represents
a simplified three-tiered Cloud with Web, Application, and Database servers. This
cloud hosts three EAs (App A, App B, App C) serving three user groups, one group per
EA. Each server corresponds to the model’s node with a number of processing units
equal to the number of CPUs in the server. The users of each EA as well as network are
modeled by dedicated nodes. All servers are the physical ones without any partitioning
among applications. Web and Application servers have 8 CPUs each; Database server
has 16 CPUs.

Figure 1 Model 1 of the Cloud hosting three enterprise applications

The models in this article were analyzed using TeamQuest solver
[http://teamquest.com/products/model/index.htm]. Workload 1 for Model 1 is characterized in
Table 1. For each application it is presented by a transaction identified by application
name. For each transaction, each user initiates it a number of times indicated in column
“Number of transaction executions per user per hour.” We analyze the model for 100, 200,
300, 400, 500, 600, 700, and 800 users.

Table 1

Workload 1 for Model 1

Number of users Number of
Transaction name Total Total Total Total Total Total Total Total Total transaction
3 100 200 300 400 500 600 700 800 executions per
user per hour
App A transaction 1 50 100 150 200 300 400 350 400 10
App B transaction 1 25 50 75 100 150 200 175 200 20
App C transaction 1 25 50 75 100 150 200 175 200 5

To solve the model we have to specify the profile of each transaction (Table 2).
Transaction profile is a set of time intervals (service demands) a transaction has spent in
all processing units it has visited while served by the application.

Table 2
Transaction Profiles (seconds)

Time in Time in Web Time in App Time in Database
Network node server node server node server node
App A transaction 0.001 0.4 2.0 10.0
App B transaction 0.0015 0.2 1.0 5.0
App C transaction 0.003 0.4 10.0 10.0

Transaction times for non-virtualized physical platform are presented in Figure 2.

Figure 2 Transaction response times for three applications

Model 1 predicts that the system can serve 600 users with acceptable transaction times.
An increase in a number of users beyond 600 causes steep transaction time growth
because the Database server reaches the maximum of its CPU capacity (Figure 3).

Figure 3 Utilization of Cloud’s servers

Figure 4 shows breakdown (by percentage) of Database server utilization by different
EAs.

Figure 4 Breakdown of utilization of Database server by applications A, B, and C

Figure 4 suggests that the largest “consumer” of Database server CPU capacity is App
A. We use this breakdown data on the next step where we divide Database server by
three VMs (one VM per application) and study how Database server virtualization
impacts performance.

Model 2 Application Performance on Virtual Platform

The virtualized Database server hosts three VMs: VM A with App A, VM B with App B,
and VM C with App C. On the Database server with 16 CPUs, one CPU accounts for
6.25% of the server’s total capacity (100% / 16 = 6.25%). A chart on Figure 4 shows the
percentage of Database server utilization by each application for 600 users:

App A: 50%
App B: 24%
App C: 12%

Based on the above data, listed below are the numbers of CPUs we have to assign to
each VM:

VM A: 50% / 6.25% = 8 CPUs

VM B: 24% / 6.25% = 4 CPUs

VM C: 12% / 6.25% = 2 CPU

Model 2 of a system with three VMs hosted on Database server is pictured in Figure 5.

Figure 5 Model 2 with virtual machines

Transaction times delivered by Cloud with virtualized Database server are presented in
Figure 6.

Figure 6 Transaction times in system with VMs having 8-4-2 CPUs

Model 2 predicts transaction time degradation for 600 users for all three EAs due to
high utilization of all three VMs (utilization exceeds 90%, Figure 7).

Figure 7 Utilization of servers in system with VMs having 8-4-2 CPUs

Model 3 Allocating additional CPUs

So far we assigned 8+4+2=14 CPUs out of 16 CPUs on the physical Database server to
three VMs. Because the largest transaction time increase happened for App C and App
A, we will distribute the remaining 2 CPUs as shown below:

VM A: 50% / 6.25% = 9 CPUs

VM B: 24% / 6.25% = 4 CPUs

VM C: 12% / 6.25% = 3 CPUs

Model 3 predicts noticeable improvement of transaction times for App C and App A
(Figure 8). As expected, adding CPUs to VMs C and A did not impact time for App B.

Figure 8 Transaction times in system with VMs having 9-4-3 CPUs

Time improvement for App C and App A was achieved by lowering CPU utilization of
VMs C and A (Figure 9).

Figure 9 Utilization of servers in system with VMs having 9-4-3 CPUs

Comparison of non-virtualized and best performing virtual deployment with VMs
having 9-4-3 CPUs is seen in Figure 10.

Figure 10 Transaction times on virtual and non-virtual platforms

Virtualized setup features longer transaction times than physical platform for all EAs
starting from 300 to 400 users. That means the capacity of all VMs (having in total the
same 16 CPUs as physical Database server) is lower than the capacity of the non-
virtualized Database server. This surprising conclusion is in line with queuing theory
analysis explaining longer transaction times in virtual systems by longer waiting times
in queues.

Quantitative consideration of that effect can be found in [Leonid Grinshpan. Solving
Enterprise Application Performance Puzzles: Queuing Models to the Rescue, Willey-IEEE
Press, 2012]. Here we offer qualitative explanation that is based on the queues behavior
in two real life environments – toll plaza and movie theater box office. A toll plaza with
booths equally accessible by any cars has lower congestion than the same plaza with
booths divided into two categories: ones serving only sedans and others serving only
trucks. An intuitive justification is: in non-divided plaza, in the absence of the trucks a
sedan can be processed by any booths and vice versa; in divided plaza the booths

dedicated to trucks will stay idle even if there is a queue of sedans. The same behavior
is exhibited by the queues in a box office – if any wicket serves any theatergoers, than
the waiting queue is not as long as in a case when some wickets provide only to
particular customer categories.

Virtualization is not a Chinese Wall

Virtualization technology isolates EAs by creating an illusion that each EA is hosted on
its own servers. But despite living in segregated environments, the applications will
impact each other’s performance when the physical capacity of a hosting platform is
overprovisioned.

We show in this paragraph how system monotoring can help to identify
overprovisioning of physical AIX platform with logical partitioning (LPAR)
[http://www.redbooks.ibm.com/abstracts/tips0119.html?Open]. LPAR is comparable to
a virtual machine in the VMware world and represents a subset of a physical server
hardware resources. Because of these similarities, the logic of our analysis is applicable
not only to AIX LPAR, but to other virtual platforms as well. To monitor a logical
partition one can use vmstat command:

vmstat t N,

where t is a sampling time interval (seconds) and N is a number of sampling intervals.
The command reports system configuration information as well as performance
counters (Figure 11):

Figure 11 System configuration and performance counters

In this example physical AIX server has 24 logical CPUs (parameter lcpu) and 31743 MB
memory (parameter mem). The command was executed for a logical partition that is
entitled to 4 CPUs (parameter ent). Entitlement value equals a number of physical CPUs
available to a partition at any time. The total number of entitled CPUs for all partitions
cannot exceed the number of physical CPUs in a partitioned server. If the partition at
any moment does not use all its entitlement, then unused capacity is transferred to the
common shared pool of CPUs and becomes available to any other partition.

CPU activity is characterised by six parameters:

us - user time percentage

sy - system time percentage

id - idle time percentage

wa - time when system processes I/O requests (percentage)

pc - number of physical CPUs consumed

ec - percentage of entitled CPU consumed (these parameters can

exceed 100%).

Figure 12 shows CPU utilization in LPAR when a hosted application consumed all four
entitled CPUs. In such a case the sum of the parameters us, sy, id and wa is 100%;
parameter pc informs that application is using almost 5 CPUs and parameter ec reports
actually the same fact but only as a percentage of an entitlement. The sum of pc
parameter readings across all LPARs is equal to the number of physical CPUs
consumed by all applications. When this sum is equal to the number of physical CPUs,
a host server is overprovisioned, affecting the performance of all hosted applications.

Figure 12 LPAR monitoring data when application demand

exceeded entitled number of CPUs

But what if the host is not overprovisioned, but a pc parameter for a particular LPAR
exceeds its entitlements? Does this mean a performance degradation of an application
in that LPAR? The answer is no; it just means that available CPUs are assigned to this
LPAR. To avoid misunderstanding follow the rule: Even if the counter’s readings look
alarming, there are no EA performance degradations as long as transaction times are
acceptable to business users.

Take away from the article

1. Queuing theory states that processing time in partitioned systems is longer than
in non-partitioned ones. Hardware virtualization conforms to that rule and that
is why we have to come to terms with the inconvenient truth that the total
capacity of all guest virtual machines is lower than the total capacity of non-
virtualized host servers for the same workloads and service demands.

2. Virtual environments inherently feature additional management overhead as
well as competition for shared components (memory bus, I/O system controller,
etc.) that slows down access to shared system resources.

3. Cloud service providers have to take into account a decrease in physical
environment capacity, and in availability of their shared components while
provisioning virtual environments for EAs.

4. Partitioning does not erect the Chinese wall among hosted applications - they
collide when the total demand for resources exceeds available physical capacity.

5. Interpretation of performance counters in virtual environments is more
sophisticated than it is in dedicated ones. To avoid misunderstanding follow the
rule: Even if the counter’s readings look alarming, there are no EA performance
degradations as long as transaction times are acceptable to business users.

About the author

During his last fifteen years as an Oracle consultant, the author was engaged hands-on
in performance tuning and sizing of enterprise applications for various corporations
(Dell, Citibank, Verizon, Clorox, Bank of America, AT&T, Best Buy, Aetna, Halliburton,
Pfizer, Astra Zeneca, Starbucks, etc.).

Enterprise application in the cloud – virtualized deployment

More Related Content

Similar to Enterprise application in the cloud – virtualized deployment (20)

Recently uploaded (20)

Enterprise application in the cloud – virtualized deployment