15
Capacty Dragmntat on Driving Eciency in Virtual & Cloud Environments Andrw Hllr, Co-Foundr & CTO, CRBA i  T e P A P e R Data Cnt r int ll gnc

CiRBA Wp Capacity ion

Embed Size (px)

Citation preview

Page 1: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 1/14

Capacty Dragmntaton

Driving Eciency in Virtual &Cloud EnvironmentsAndrw Hllr, Co-Foundr & CTO, CRBA

WH

i  T e P A P e R 

Data Cntr intllgnc

Page 2: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 2/14

Page 3: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 3/14

1.866.731.0090 3

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

introducton

A major shit is occurring in the way capacity is managed in the data center. Previously,

organizations added capacity to IT environments by installing a new server, installing theapplication to be hosted on it, and occasionally perorming upgrades or server “reresh” in

order to avoid supportability issues. This “one app per box” mentality was a very rigid model,

and unortunately has created a legacy o excess compute capacity in most data centers.

 The rise o X86-based virtualization technologies like VMware has changed the way supply

and demand should be viewed. The notion o “placement” and the ability to combine

multiple workloads onto a single host is a disruptive element in data center and capacity

management. Virtualization changes the way in which environments are planned and the

tooling required to manage them.

  Traditional capacity management oriented rameworks do not contemplate the notion

o “placements.” Rather, their ocus has traditionally been to track utilization levels inorder to plan uture capacity and avoid “saturation”. Unortunately, the models used or

this assume a rigid environment where workloads are not mobile, and they oten break 

down i a workload is relocated to a dierent server. In other words, because they do not

understand the notion o placements they cannot determine how to optimize them.

And because it is not possible to proactively optimize the eciency o a modern data

center without optimizing workload placements, a signicant gap has been created in the

capacity management world.

 The importance o placements, and the impact o not optimizing them, can be illustrated

by making an analogy to the game o Tetris. Tetris is a game o placement, and embodies

many o the undamental concepts o supply and demand that are observed when

optimizing workload placement in virtual or cloud-based IT environments

Page 4: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 4/14

1.866.731.0090 4

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

Th Ttrs efct

Finding the Optimal DensityWhen planning and managing virtual environments, one o the rst questions that is asked

is how much ts on a given server? This is an important question, as an environment’s

“density” dictates how many servers are required, how many hypervisor licenses are

needed, and generally has a huge impact on inrastructure costs. Erring on the side o low

density can increase costs signicantly, while overly high density tends to introduce risks,

including SLA penalties and even the risk o ailure.

It is in optimizing this density that the Tetris analogy becomes relevant. Imagine that

available server capacity like the blocks, is analogous to the Tetris playing eld, and the

workload demands are analogous to the dierent Tetris blocks. Workloads have dierent

characteristics sizes and shapes. Also how they t together is important when you start

talking about multi-tenancy models such as virtualization and cloud computing, where

optimizing resource sharing and over-commit become critical design points.

Page 5: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 5/14

1.866.731.0090 5

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

The Myth o the “Average Workload” To look at this in simplistic terms, we will rst look only at averages. In our Tetris analogy,

we will assume that each individual block is on average 4 units in size, and the playing eld

happens to be 84 units. That would imply that 21 o the blocks would t in that playing

eld. Unortunately it is not that simple. Even when using Tetris blocks, which are designed

to t together perectly in the right combination, you will typically not be able to make use

o the complete playing eld. This is because the sizes and shapes heavily impact how the

blocks t together, making ecient use o capacity very dicult.

 The odds are clearly stacked against us when we look at it this way – we cannot simply look 

at the total work that is being done and the total capacity that is available to determine

how best to place workloads. In a virtual environment, as in Tetris, optimal placements

really depend on a lot more than just the averages.

Page 6: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 6/14

1.866.731.0090 6

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

Capacity Deragmentation The rst step in minimizing stranded capacity is to not use averages or planning purposes.

Instead, it is very important to look at the patterns or “shapes” o workloads. Analyzing

workload patterns, which are analogous to the shapes o the Tetris blocks, is critical to

eciently combining workloads. For instance, it is critical to think about a workload as it

changes during an operational cycle, and the pattern that it carves out throughout the

day, month, and even the entire year. Moreover, the eectiveness o resource over-commit,

which is a undamental capability o virtualization, is heavily dependent on dierent

consumers using those resources at dierent times. I workloads utilize the same resources

at the same times, the over-commit model will create a lot o risk. But i workloads are

placed in a way that makes them “dovetail”, then these risks are avoided, in much the same

way that Tetris blocks are manipulated in order to make them t together.

  To plan in this more sophisticated way requires more than just a simplistic “trend and

threshold” model, and instead requires consideration o the permutations and combinations

o placements that will produce the optimal outcome. By assessing dierent variations o 

placements (o which there can be many) it is possible to nd the solution that gives the

highest eciency and uses the least space, while incurring the lowest risk.

Page 7: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 7/14

1.866.731.0090 7

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

Applyng Ths to th Ral World

 The challenge in all o this is that Tetris blocks are designed to t together. Workloads are

much more complex. Workloads are not a xed shape, they actually carve out dierentshapes and sizes throughout the day. Also, each workload is unique, and within a virtual

environment there can literally be billions o dierent ways that they can be placed

together. For instance, in moving 15 workloads onto 5 servers, there are approximately 40

billion dierent placement combinations. This underscores the need to use permutations

to understand which one o those combinations is the best one.

Constraints on Workload PlacementsFurther complicating this is the act that there are many constraints that impact the

placement o workloads in virtual and cloud environments. This means that o the

billions o potential combinations, a large portion will be invalid or inappropriate

due to practical constraints on the environment. These constraints include technical

considerations (platorm compatibility, hypervisor compatibility, connectivity, etc.),

business considerations (criticality, data sensitivity, process, security, compliance, etc.),

and resource limitations (CPU, memory, I/O, etc.).

 These constraints tend to have one o several eects: they will either rule out placement

in certain environments (e.g. sensitive data cannot go into external public clouds), create

anities between workloads (e.g. chatty applications should be placed together), or create

anti-anities (e.g. application cluster components should be kept apart). Ultimately, i the

placement combinations are not evaluated in this light, an environment cannot be optimized,

as the resulting solution may violate the undamental hosting rules in an environment.

Conguration

Optimal Answers

UtilizationBusiness

Page 8: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 8/14

1.866.731.0090 8

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

Multi-Dimensional Analysis Model or Optimizing Placements and Allocations

Optimizing Workload PlacementsIn order to solve this ormidable combinatorial challenge, it is necessary to decompose

the problem into its undamental components and solve the resulting multi-dimensional

problem. This approach is ideally based on empirical data that includes accurate

conguration inormation, up-to-date business inormation, and historical utilization data.

By assessing these parameters against rules and algorithms that capture the preerences

and constraints in a given environment, it is possible to solve or the optimal workload

placements and resource allocations. When compared to traditional capacity management

approaches, this is somewhat o a paradigm shit, as it is undamentally dierent rom the

trend-and-threshold approaches used in physical environments.

Page 9: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 9/14

Page 10: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 10/14

1.866.731.0090 10

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

I we analyze the 15 workloads and place them onto the target server platorm using this

approach we see that 4 servers are needed to host them:

I we look closely at this result, however, we see that the aggregate utilization o the

environment is not very good, which is a common symptom o capacity ragmentation.

Drilling down into the utilization levels o each o the 4 servers gives insight into why

this is:

Source Servers Stacked on Target - Like Times (Avg 20.79%)

Source Servers Stacked on Target - Like Times (Avg 16.71%)

Source Servers Stacked on Target - Like Times (Avg 6.95%)

Source Servers Stacked on Target - Like Times (Avg 4.85%)

Page 11: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 11/14

1.866.731.0090 11

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

I we examine the patterns o each o the our servers we see that each reaches a high

level o utilization at some point during the day, and that adding another workload that

requires resources at that time is not possible. The problem, however, is that the two

servers on the let are peaking in the morning, and the two on the right in the evening,

which is intuitively sub-optimal. This is analogous to stacking Tetris blocks up one side o the playing eld until it is ull, and then starting over and stacking up the other side.

Enough said.

Although this is clearly sub-optimal, it is extremely common in the data center, and can be

very dicult to detect and correct, particularly when hundreds or thousands o workloads

are involved. Many organizations wonder why their utilization levels are not higher, but

ail to consider this eect.

Deragmentation-Based Approach

  To do this properly, a permutation-based approach is required. This ensures that

ragmentation is minimized by determining the precise combinations o placements that

make best use o available capacity. And although data center workloads are not as tidy as

 Tetris blocks, the results are equally dramatic (and no less satisying).

 To illustrate, we will analyze the same workloads onto the same target platorm, but will

employ a permutation-based algorithm that assess each relevant combination against the

overall optimization criteria (as well as all technical, business, and resource constraints):

Page 12: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 12/14

Page 13: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 13/14

1.866.731.0090 13

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

Copyright © 2004-2010, CiRBA Inc. All Rights Reserved www.CiRBA.com

Concluson

Capacity deragmentation is a concept that is becoming increasingly important in the

management o modern data centers. As virtualization increases its penetration intoproduction environments, and as public and private clouds move to the oreront o the

IT mindset, the ability to leverage this newly-ound agility while at the same driving high

eciency (and low risk) is a real game changer. Only by looking at the problem rom new

and innovative angles, however, can the managers o IT environments make the transition

rom old-school capacity management to new-school eciency management.

About the Author

Andrew Hillier, Co-ounder & CTO, CiRBA, Inc.

Andrew Hillier has over 20 years o experience in the creation and

implementation o mission-critical sotware or the world’s largestnancial institutions and utilities. A co-ounder o CiRBA, he leads

product strategy and denes the overall technology roadmap or

the company.

Prior to CiRBA, Mr. Hillier pioneered a state o the art systems

management solution which was acquired by Sun Microsystems

and now serves as the oundation o their fagship systems

management product, Sun Management Center. Mr. Hillier has also

led the development o solutions or major nancial institutions, including xed income,

equity, utures & options and interest rate derivatives trading systems, as well as in the

elds o covert military surveillance, advanced trac and train control, and the robotic

inspection and repair o nuclear reactors.

Mr. Hillier holds a Bachelor o Science degree in computer engineering rom The University

o New Brunswick.

About CiRBACiRBA Data Center Intelligence (DCI) sotware enables organizations to saely maximize

eciency through the intelligent planning and management o capacity within physical and

virtual inrastructure. Only CiRBA’s policy-driven analytics accurately answer the questions

o how much capacity is truly required, how it should be allocated, and where to place

workloads in order to maximize utilization. For more inormation, visit www.cirba.com.

Page 14: CiRBA Wp Capacity ion

8/8/2019 CiRBA Wp Capacity ion

http://slidepdf.com/reader/full/cirba-wp-capacity-ion 14/14

 C  a  p a  c i    t    y  D  ef   r  a  gm en t   a  t  i    on

45 Vogell Road, Suite 600

Richmond Hill, ONCanada, L4B 3P6

Toll Fr: +1.866.731.0090

Tlphon: +1.905.731.0090

Fax: +1.905.770.4082

Onln: www.cirba.com

Copyright © 2010, CiRBA Inc. All rights reserved.

Capacity Deragmentation