Page 1
© 2014 Amazon.com, Inc. and its affiliates. All rights reserved. May not be copied, modified, or distributed in whole or in part without the express consent of Amazon.com, Inc.
More Nines for Your Dimes: Improving Availability and Lowering Costs
using Auto Scaling
Derek Chiles, AWS Solutions Architecture
(@derekchiles)
July 10, 2014
Page 2
Topics We’ll Cover Today• Auto Scaling introduction
• Console demo
• Maintaining application response times and fleet utilization
• Handling cyclical demand, unexpected “weather events”
• Auto Scaling for 99.9% Uptime
• Single-instance groups
• Cost control and asymmetric scaling responses
• CloudFormation, custom scripts, and multiple inputs
• Using performance testing to choose scaling strategies
• Dealing with bouncy or steep curves
AWS
The Weather Channel
Nokia
Adobe
Dreambox
Page 3
Ways You Can Use Auto Scaling
Launch EC2 instances
and groups from
reusable templates
Scale up and down as
needed automatically
Auto-replace
Instances and
maintain EC2 capacity
Page 4
Common Scenarios
• Schedule a one-time scale out and flip to production
• Follow daily, weekly, or monthly cycles
• Provision capacity dynamically by scaling on CPU, memory,
request rate, queue depth, users, etc.
• Auto-tag instances with cost center, project, version, stage
• Auto-replace instances that fail ELB or EC2 checks
• Auto-balance instances across multiple zones.
Prepare for a Big Launch
Fit Capacity to Demand
Be Ready for Spikes
Simplify Cost Allocation
Maintain Stable Capacity
Go Multi-AZ
Page 5
Demo
Learn the new terms:
Launch Configuration
Auto Scaling Group
Scaling Policy
Amazon CloudWatch Alarm
Amazon SNS Notification
Page 7
What’s New in Auto Scaling
Better integration
• EC2 console support
• Scheduled scaling policies in
CloudFormation templates
• ELB connection draining
• Auto-assign public IPs in VPC
• Spot + Auto Scaling
More APIs
• Create groups based on running
instances
• Create launch configurations based
on running instances
• Attach running instances to a group
• Describe account limits for groups
and launch configs
Page 8
Why Auto Scaling?
Scale Up Control CostsImprove Availability
Page 9
Why Auto Scaling?
Scale Up Control CostsImprove Availability
Page 10
The Weather Company
• Top 30 web property in the U.S.• 2nd most viewed television
channel in the U.S.• 85% of U.S. airlines depend on
our forecasts• Major retailers base marketing
spend and store displays based on our forecasts
• 163 million unique visitors across TV and web
Page 11
Wunderground Radar and
Maps
100 million hits a day
One Billion data points per day
Migrated real-time radar mapping system wunderground.com to
AWS Cloud
Page 12
30,000Personal
Weather
StationsSource: Wunderground, Inc. 2013
Page 13
Why Auto Scaling?
Page 14
Why Auto Scaling?
Page 15
Why Auto Scaling?
Page 16
Why Auto Scaling?
Page 17
Why Auto Scaling?
Hurricane Sandy
Page 18
Before Migration – Traditional IT Model doesn’t scale well
Server Count(110 Servers)
Avg. CPU Load HTTP Response Latency(~6000 ms)
HTTP Response Latency(5-15ms)
Server Count(from 110 to 170 Instances)
Avg. CPU Load
After Migration - Wunderground Radar App
Page 19
Radar on AWS Auto Scaling Architecture
Page 20
Radar on AWS
CPU Utilization
Page 21
Radar on AWS
Host Count
Page 25
Scale up to ensure consistent
performance during high-demand
Page 26
Why Auto Scaling?
Scale Up Control CostsImprove Availability
Page 27
Auto Scaling
for 99.9%
Uptime
Page 28
Here.com Local Search Application
• Local Search app
• First customer facing
application on AWS
• Obvious need for
Uptime
Page 29
Here.com Local Search Architecture
US-East-1
US-West-2
EU-West-1
US-East-1a
Zookeeper1
Zookeeper2
Zookeeper3
Frontend
Group
Backend
Groups
US-East-1b
Zookeeper1
Zookeeper2
Zookeeper3
Frontend Group
Backend Groups
AP-Southeast-1
Page 30
Here.com Local Search Architecture
US-East-1
US-West-2
EU-West-1
US-East-1a
Zookeeper1
Zookeeper2
Zookeeper3
Frontend
Group
Backend
Groups
US-East-1b
Zookeeper1
Zookeeper2
Zookeeper3
Frontend Group
Backend Groups
AP-Southeast-1
Single-Instance Auto Scaling
Groups (Zookeeper)
1. Auto-healing: Instances auto-register in
DNS via Route53
2. Dynamic: Auto Scaling Group Names
are used for cluster-node lookups
(cluster1-zookeeper1)
3. Used Standard Tools such as DNS
instead of Queries or Elastic IPs
Page 31
Auto Scaling when upgrading
without any downtime
Page 32
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
Page 33
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
New Data
V2
Page 34
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
New Data
V2
Page 35
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
New Data
V2
Page 36
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
New Data
V2
Page 37
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
New Data
V2
New
v2New
V2
Page 38
Map Data on S3
US-East-1a
Zookeeper1
cluster1
old old
New Data
V2
New
v2New
V2
Page 39
“Auto scaling”
Web Server Fleet(Amazon EC2)
Database Fleet(RDS or DB on EC2)
Load Balancing(ELB)
v1.1 v1.1
v1.1 v1.1
v1.2
v1.2
v1.2
v1.2
Auto scalingMax instances
Min instances
Scaling Trigger
Custom Metrics
Upper Threshold
Lower Threshold
Increment by
Common scenario: Blue Green Deployments
Using Auto Scaling
Page 40
Here.com Local Search Success
• Increased Uptime to 99.9%
• All detected health problems have been successfully replaced by Auto Scaling with zero intervention.
• Zookeeper setup has performed flawlessly
“We’ve been
paranoid so it still
pages us; It’s
beginning to feel
silly.”
Page 41
Why Auto Scaling?
Scale Up Control CostsImprove Availability
Page 42
Adobe Creative
Cloud Runs on
AWS
Page 43
Adobe Shared
Cloud Architecture
on AWS
Page 44
Auto Scaling the Web Layer
Based on Number of HTTP requestsAverage CPU loadNetwork in/out
Page 45
Auto Scaling the Web Layer
Auto Scaling the Worker Layer
Based on SQS queue length
Based on Number of HTTP requestsAverage CPU loadNetwork in/out
Page 46
Scale up fast, scale down slow
Page 47
Cost Control
• Scheduled scaling: we analyzed our traffic and
picked numbers. – scale up in the morning, scale down in the evening
• Policies for slow scale down
• Stage environments: downscale everything to
“min-size” daily (or more)
Page 48
CloudFormation + Auto Scaling"ScaleUpPolicy" : {
"Type" : "AWS::Auto Scaling::ScalingPolicy", "Properties" : {
"AdjustmentType" : "ChangeInCapacity", "Auto ScalingGroupName" : { "Ref" : "WorkerAuto ScalingGroup" }, "Cooldown" : {"Ref": "cooldown"}, "ScalingAdjustment" : { "Ref" : "adjustup" }
} },"WorkerAlarmScaleUp": {
"Type": "AWS::CloudWatch::Alarm", "Properties": {
"EvaluationPeriods":{"Ref" : "evalperiod"}, "Statistic": "Sum", "Threshold": {"Ref" : "upthreshold"}, "AlarmDescription": "Scale up if the work load of transcode queue is high", "Period": {"Ref" : "period"}, "AlarmActions": [ { "Ref": "ScaleUpPolicy" }, { "Ref" : "scalingSNStopic" } ], "Namespace": "AWS/SQS", "Dimensions": [ { "Name": "QueueName", "Value": {"Ref" : "queuename" }}], "ComparisonOperator": "GreaterThanThreshold", "MetricName": "ApproximateNumberOfMessagesVisible"
Page 49
How – Custom Metrics. . .
Sat Oct 6 05:51:03 UTC 2012
Number of AZs: 4
Number of Web Servers: 16
Number of Healthy Web Servers: 16
ELB Request Count: 9523.0
Request Count Per Healthy Web Server: 595.1875
Network In Per Healthy Web Server: 51 MB
Network Out Per Healthy Web Server: 1 MB
CPU Per Healthy Web Server: 25.23875
Publishing Custom Metrics: InstanceRequestCount, HealthyWebServers, InstanceNetworkIn, InstanceNetworkOut, InstanceCPUUtilization to namespace WebServer in us-east-1
. . .
Page 50
How – multi-input scaling
Scale up
Scale down
+2 instances if more than 50 visible messages for >5 min
+50% instances if more than 1000 msg for >2 min
+ fixed 100 instances if more than 10000 msg for >1 min
-10 instance if 0 msg for more than 10 min
-25% if 0 msg for more than 30 min
Page 51
Adobe’s Advice
• Use CloudFormation!
• Know your system, thresholds
• Watch your scaling history
• Scaling up is easy, scaling down not so much
• Mantra: scale up fast; scale down slow
Page 53
Scaling strategies we use
Scaling with
CloudWatch alarmsScheduled scaling (onetime, recurring)
Page 54
A little background on our application
• Ruby on Rails
• Unicorn
• We teach kids math!
Page 55
A workload well suited for auto scaling
Page 56
Scaling with CloudWatch alarms
Page 57
Performance test to get a baseline• Discover the ideal number of
worker processes per server– Too few and resources go
unused
– Too many and performance suffers under load
• Obtain the maximum load sustainable per server– Our performance tests
measures number of concurrent users
• Find the chokepoint– For us, this was CPU utilization
Page 58
Performance testing
Page 59
Identify the breaking point
Breaking point was at about 400 users per server
Page 60
Our first method to find scale points• Provision a static amount
of servers that we know can handle peak load
• Adjust scale up and scale down alarms based on observed highs and lows
• This worked, but was super inefficient, both in time and money spent
Page 61
Let’s do some math – identify variables
Independent
• Concurrent users
Dependent
• CPU utilization
• Memory utilization
• Disk I/O
• Network I/O
Page 62
Let’s do some math – find the slope• Adding about 1600 users per hour
• Which is about 27 per minute
• We know that we can handle a max of
about 400 users per server at 80% CPU
usage
• Which is about 0.2% CPU usage per user
Page 63
Let’s do some math – when to scale?
• We know (from other testing) that it takes
us about 5 minutes for a new node to
come online
• We’re adding 27 users per minute
• Which means we need to start spinning
up new nodes when we’re about 135
users ( 27 x 5 ) per node short of max
• Which is at about 53% utilization:
(80% - (0.2% * 135))
Page 64
How much to scale up by?
• The lowest we can scale up by is 1 node per AZ,
otherwise we would be unbalanced
• For us, this is an extra 800 users of capacity in
five minutes, plenty enough to keep up with our
rate of adding 1600 users per hour
• Adding 800 users of capacity every five minutes,
we could support 9600 additional users per hour
Page 65
Evaluate your predictions• In the real world, we’ve inched up from
scaling at 53%
• Our perf test is a little harsher than the
real world
• Numbers derived from the perf test are
only as accurate as the simulation of
traffic you in your perf test.
Page 66
Scheduled scaling
Page 67
Acceleration in load is not constantRequest count for a 24 hour period
Page 68
We can’t use one size fits all• Scale too aggressively
– Overprovisioning: increases
cost
– Bounciness: we add more
than we need and have to
partially scale back shortly
after scaling up, which
increases cost
• Scale too timidly
– Poor performance
– Outages due to lack of
capacity
Page 69
Putting it all together
Page 70
The opportunity cost of NOT scaling
• Our usage curve
from 3/20
• Low of about 5
concurrent users
• High of about
10,000 concurrent
users
Page 71
The opportunity cost of NOT scaling
• No autoscaling
• 672 instance hours
• $302.40 at on-
demand prices
Page 72
The opportunity cost of NOT scaling
• Autoscaling four
times per day
• 360 instance hours
• $162 at on-
demand prices
• 46% savings vs no
autoscaling
Page 73
The opportunity cost of NOT scaling
• Autoscaling as
needed, twelve
times per day
• 272 instance hours
• $122.40 at on-
demand prices
• 24% savings vs
scaling 4 times per
day
• 60% savings vs no
autoscaling
Page 74
The opportunity cost of NOT scaling
$302/day
$162/day
$122/day
Page 75
Demand curve hugs the usage curve…
Page 76
…and a (mostly) flat response curve
Page 77
“Auto Scaling saves us a lot of money; with
a little bit of math, flexibility of AWS allows
us to further save by aligning our demand
curve with usage curve.” -- Dreambox
Page 78
Why Auto Scaling?
Scale Up Control CostsImprove Availability
Page 79
Key Takeaways
• Maintaining application response times and fleet utilization
• Scaling up and handling unexpected “weather events”
• Auto Scaling for 99.9% Uptime
• Single-instance groups
• Cost control and asymmetric scaling responses
• CloudFormation, custom scripts, and multiple inputs
• Using performance testing to choose scaling strategies
• Dealing with bouncy or steep curves
The Weather Channel
Nokia
Adobe
Dreambox
Page 80
Thank You!
Derek [email protected]
@derekchiles