Date post: | 09-Jun-2015 |
Category: |
Documents |
Upload: | hp-software-solutions |
View: | 1,186 times |
Download: | 4 times |
©2010 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice
Session ID: BTOT-TU-1700/9Twitter hashtag #HPSWU
©2010 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice
Speaker Name: Shawn Islam, Hari KannanDate: Nov 30, 2010Session ID: BTOT-TU-1700/9
Capacity Management and Resource Optimization in Cloud* environmentsTechniques for Identifying Optimization Opportunities in a Virtualized Cloud
* Virtualized cloud environments
3
A Balancing ActCapacity Management
“A discipline that ensures IT infrastructure is provided at the right time in the right volume at the right price, and ensuring that IT is used in the most efficient manner”
- ITIL definition
Meeting IT Supply with Business Service Demand “Just-in-Time”
“The growing adoption of virtualization (and related technologies, such as cloud computing), plus changing organizational and process demands, will force a reassessment of traditional capacity planning and related IT planning functions. ...Traditional capacity planning tools use a silo-based approach to analytical modeling to determine capacity. Such an approach will be inadequate because these tools must be able to build an end-to-end model that is based on IT and Business Services...”
Source: Gartner, IT Resource Planning: Going Beyond Capacity Planning, Cameron Haight, Milind Govekar, George J. Weiss , February 2009
Capacity Management: Before Virtualized Cloud
Typically, each server hosted a single workload Very limited or virtually no impact of one service’s demands on another service
Capacity statically configured for peak workloads Results in low average utilization
Deploying additional capacity time consuming Time is €€€ !!!
Not possible to migrate workload without downtime Results in a more stable, predictable capacity demand
Capacity management was silo-based Lack of holistic view resulted in inefficiencies
Primary focus on core infrastructure elements (CPU, memory, IO) End-User not part of the capacity equation5
6
Virtualization adoption is spiraling
2008 2009 20120
10
20
30
40
50
60
5.80 10.80
55.00
VM Installed Base (Millions)
Avg VM Density: 8
Avg VM Density: 75
Capacity Management & Planning is the No.1 challenge
facing companies as they virtualize, according to Forrester
“Admins are aiming for 70% or greater physical server resource utilization by 2012”, according to IDC
2008 2009 2010 2011 201205
101520253035404550
1219
2838
48
Percentage of Installed x86 Work-loads Running in a VM
Sources: Gartner, Forrester, IDC
Virtual Machine installed base to grow by 5x in three years
Almost half of all workloads will bedeployed in virtual environments by
2012
7
CLOUD: The BIG picture
Cloud Storage ComputeResources.
NetworkBandwidth
Power
DB
Java Apps
IT and Helpdes
k
Operations Automatio
n
Operations Monitoring
Capacity Planning
and Optimizatio
n
Capacity Management in a Virtualized Cloud
Multiple workloads (VMs) per host Results in higher average utilization, but high impact of one service’s
demands on another service
Capacity easily scaled up or down Deploying additional capacity fast and cheap Unprecedented volume and frequency of workload deployment can result in VM Sprawl High need for tracking and reporting capacity and usage metrics
Dynamic workload migration without downtime Can result in unexpected load spikes on server and network High need for headroom-based capacity provisioning
Capacity shared across multiple tenants & Business
Services Need for Capacity Management and Reporting according to Business needs
8
Workload Placement: Challenges in a Virtualized Cloud
As VM density increases, sizing and workload placement becomes complex -
need to understand the impact of resource sharing
Using individual VM’s peak or average will result in sub-optimal performance
In-house Tools (think “Excel”) don’t scale well, need constant tune-up
Remaining “head-room” capacity for unforeseen spikes in demand unknown
Why workload placement is important?
10
10 workloads, 20% Average Utilization
Hosted on 2 servers
10 workloads, 80% Peak Utilization
Using Averages leads to under-provisioningUsing Peaks leads to over-provisioning
Hosted on 8 servers
Why workload placement is important?
– What if we knew that there are 2 types of VMs?• VMs whose CPUs peaks during day and VMs whose CPUs peak at night
11
10 workloads5 VMs peak @ 80% during
day5 VMs peak @ 80% during
night
Hosted on 5 Servers
Using seasonality data leads to optimal provisioning
Capacity Management – The Complexity–Now, consider multiple dimensions
12
Compare•HypervisorsVMware,Hyper-V
•CPU - AMD, Intel
•H/W models
Analyze•CPU
•Memory
•I/O
•Energy
Con
stra
ints
•B
usin
ess
•Te
ch
nic
al
Excel no longer adequate!
13
– Clear understanding of Virtual resource dependencies to Application or Business Service: Workload to Application Service mapping, identify shared workloads
– Order application workloads from largest demand to smallest for different time intervals• Sum up the time varying traces for each resource instance across multiple dimensions to get total per-interval
demand
• Peak of sums is typically less than sum of peaks (15 vs. 19)
– Order resource instances from greatest capacity to smallest
– Across all time intervals, simulate by doing a first fit by adding resources
– Repeat this using randomness to provide many different possible initial solutions
– Converge on the best-fit solution
An Algorithm driven approach for Workload Placement
Time
App1 App2 Total
0 10 5 15
1 3 4 7
2 2 9 11
14
Analytics: What statistical measure to use– Using peak values and using average values have their limitations
• Estimates can lead to over-sizing and under-sizing
– Trend analysis is required in order to arrive at accurate estimates• percentile values (95-percentile) and sustained peaks rather than the absolutes help to
eliminate short-term and infrequent spikes
• Time varying historical trends needs to be analyzed for each resource instance to identify the total average and peak demand of different time intervals
• Future planned growth and forecasting
– Manual sizing is impossible for large number of workloads: need analytics engine to carry out the algorithm
– What if analytics for balancing Resources and Demand• For all time interval satisfy 100% demand vs. less than 100% demand is sufficient for some time intervals
• Quality of Service needs to be considered for determining demand accommodation and head rooms
Constraints and Headroom setting best practices
– Place apart• VMs running workloads that perform lot of disk IO operations
• VMs that are members of a failover cluster (MSCS, Veritas, etc)
• Two applications must not share same resources
– Place together• In case of VMware, place VMs with similar workloads running same application versions on same host to gain
from TPS*
• Application bound to use a resource group
– Headroom (free CPU cycles, memory) must be sufficient to allow sudden peak demand, and for free migration of VMs across hosts in a cluster – aim for 70% utilization rates• Provide headroom for network usage too
• Factor in VM migrations also (Ensure that there is sufficient free capacity in the cluster available for free migrations)
• After failover utilization rates at cluster level must not exceed 80%.
• Ensure that there are a minimum of 4 paths from host to storage to allow for any fiber connection failures
15
16
Analysis & Visibility needed for Decision Planning
Max
90%AVG
Usage - Trend
Inventory – Data Center
Business App to Infra Inventory
Forecast
Forecast30, 60, 90 days
AVGTrend
Capacity
17
– DRS is reactive capacity Management, pro-active capacity management complements DRS and can reduce the need for live migration
– DRS/HA does not account for future planned growth
– DRS is CPU/memory focused, holistic capacity management adds value
– DRS is silo-based (restricted to a single cluster), datacenter-wide capacity management is a must
Capacity Management in context of DRS/HA
VMware HA
Cluster
VMware DRS
Cluster
Are both really needed? {YES}
18
– Faults and incidents from event management systems act as triggers for capacity analysis• Right-sizing VMs
• Fitting VMs to hosts with unused capacity
• Cloud-burst – can take services temporarily from cloud providers
− Ensure that security constraints are met at application/user level
– Business Services and IT get linked via the Runtime Service Model• Visibility into Service Performance
• Easy to determine capacity needs and assign resources for new initiatives
Capacity Management converged with Service and Operations Bridge
ReviewLet’s summarize
Cloud Capacity Management: Key Requirements
20
VISUALIZE
What you have, What you use, What can fit, What can be
improved
OPTIMIZE
Multi-Dimensional, Time-Varying Analysis and
Workload Placement
recommendation•Support Heterogeneous Environments•Scalable across 1000’s of nodes•Easy integration with existing Data Collectors•Continuous, Real-Time Monitoring and Analysis•Visibility into Service Performance•Correlate Infrastructure Performance with Business Service Performance
FORECAST AND PLAN
What-if Scenarios, Trend
Analysis
Capacity Management: Key Questions• How much Capacity do I have and how much do I use?
• Across datacenter, by Business Service, by cluster• Across Physical, VMware, Hyper-V resources
• Does my infrastructure keep up with end-user response?• Do I have idle and powered-off VMs resulting in a VM sprawl?• Which servers are under or over-configured?
• How much room for growth do I have?• Across datacenter, by Business Service, by cluster
• When will I run out of capacity?• What-if I add or remove workloads?• What-if I add/modify/remove servers, memory, or CPU?• Should I buy server model A or model B?• What if Business grows by 5% each quarter?• I want to increase my response time, how much more resources
do I need?
Forecast and Plan
• How do I place my VMs optimally in a server farm?• Are my VMs sized correctly?• What is the right VM size for my physical server?• Do I have enough headroom for unexpected spikes?• Will my new workload fit in my existing cluster?
Optimize
Visualize
21
Continue the conversation with your peers at the HP Software Community hp.com/go/swcommunity