Application Profiling for Memory and Performance

Application Profiling
for
Memory & Performance
Srinath Perera Pradeep Fernando Dinuka Malalanayke
WSO2 Inc WSO2 Inc WSO2 Inc

Understanding Server Performance

When Concurrency Increases?
• Throughput increases, but where will it stop?
• When there are contention for the server
• Contention for what?
o CPU
o Memory
o Disk - I/O
o Network -I/O
• If your server running on full capacity, there should be at
least one resource under contention.

How to measure Contention?

• CPU - CPU or load average
• Memory - GC logs, monitor memory, sometime swapping
• Disk I/O - IOPS per seconds, load average
• Network I/O - network utilization, load average
• What is load average? length of processor queue of the OS
• How much load average is too much?
o Load average = number of cores (machine is fully used)
o Load average = 2* number of cores (machine is loaded)

If you cannot find contention
• Either load is too small
• You have too much locks. Look for lock profile
• Server might be processing with too less threads (IO and
CPU thread switches)

Performance Tuning
• Increase throughput/ latency or both
• Know that latency vs. throughput is often a tradeoff
• Run the system and see how it does, make sure you are
putting in enough load
• Verify where is the bottleneck? is it CPU, memory, network,
or disk
• Start tuning

Tuning

• When you tune, bottlenecks will shift, you will have to shift
your focus
• Can you tune settings (Application, JVM, OS) to improve and
shift the bottleneck? play with parameters .. we call this
parameter sweep
• Then focus on the code, look for right profile
o CPU profile
o Memory/ Allocation profile
o Network profile
o IO profile
o Database profile

Programmer Nightmares

ClassCastException
Permgen Errors
DeadLocks
NullPointer
Unacceptable Latency
CPU spin OutofMemoryException
ArrayIndexOutOfBoundsException

[image] http://www1.free-clipart.net/gallery2/clipart/Holidays/Halloween/Ghost_-_Boo.jpg

Introducing: Profiling...

• Best thing would be to write bug free, high performant code
• Normally development happens in iterative manner - get
something to work and improve on that.
• Basic profiling can help you to eliminate most of the memory
and performance issues.

Right Tools...
• There are many commercial and open source tools
• We are going to use,
o Jprofiler by ejTechnologies
 CPU, Memory, Allocation, Network, DB
o Eclipse Memory Analyzer tool
 Find memory leaks
o Standard JDK tooling - Jconsole
 Basic stats

[image] http://writingsongs.com/pictures/toolbox.jpg

Environment Setup
With WSO2 Carbon

Setting up JProfiler
• Profiling modes local/remote
• Applying common filters for WSO2 products
• Different views..

JProfiler Contd.
CPU view
- Call trees, Hot spots
Memory view
- Allocation Hot spots, Recorded objects, Heap walker
VM telemetry view
- Memory, Threads status graphs
Probes
- JDBC , Sockets

JConsole
• Connect to the running instance local/remote
• Different memory banks
• Getting most out of Mbeans.
• Getting a memory dump and analyzing with Eclipse MAT.
• Remote JMX URL gets printed in the console during server
startup:

JMX Service URL : service:jmx:rmi://localhost:11111/jndi/rmi://localhost:9999/jmxrmi
username : admin password : admin

Memory Analyzer Tool
• Analyze heap dumps.
• Identifying OOM issues are trivial with basic leak report.

Usual Suspects

Some of the common issues...

Permgen Errors
• Permgen space relates to the 'Programme code' part, a.k.a -
Loaded classes.
• ClassLoader leaks
• Typically happens in Container environments
• Can uncover by doing deploy/redeploy cycles.
• If one of the application objects get referenced by an object,
outside the application, permgen error waiting to happen.

Permgen Errors

Container

URLClassLoader Application

Application Objects
Container provided
Objects

Memory Leaks - OOM
• Languages like C, gave the control of memory to the
programmer.
• Java is a managed memory, language.
• Still the JVM can't free up the memory, if the application is
holding on to the objects.
• Unnecessary accumulation of objects,
o Object creation for each and every connection
o Slow output rate in the input/output system
o Collecting clusterwide messages without an upper bound.

DeadLocks

Resource A

Resource B

• Re-designing the lock acquiring sequence
• Using LockManagers that keeps track of lock acquisition

Liveness issues...

• Starvation - A thread can starve for CPU cycles, if it is
blocked by a lock-wait.
• Poor Responsiveness
• LiveLock scenarios

Programming Best Practices

• Think about memory aspects/ data retrieval process while
writing code.
• Optimize the critical Path (look at CPU profile).
• Make use of Standard libs as much as possible.
• Profile your apps before releasing them to production.

Identity Server - Losing weight

• Carbon server optimized to run in constrained memory
environment.
• No front end components - 'createWorker' ant task.
• Selective Admin service loading. -Doptimized=true
• Removed unwanted functionality based on the use-case,
(p2 - provisioning, etc)

Raspberry-Pi Clusters

• Application server
cluster running on R-
Pis.
• Powering the
WSO2Con 2013
mobile app
• 512MB of RAM and
ARM processor.

Demo

CPU spin and Memory leak scenario

Application Profiling for Memory and Performance

More Related Content

What's hot (20)

Similar to Application Profiling for Memory and Performance (20)

More from WSO2 (20)

Application Profiling for Memory and Performance