Transcript Optimizing your Java Applications for multi

Optimizing your Java Applications
for multi-core hardware
Prashanth K Nageshappa
[email protected]
Java Technologies
IBM
1
Agenda
• Evolution of Processor Architecture
• Why should I care?
• Think about Scalability
• How to exploit
• Parallelism in Java
• JVM optimizations for multi-core scalability
2
As The World Gets Smarter, Demands On IT Will Grow
Smart supply Intelligent
chains
oil field
technologies
Smart food
systems
Smart
healthcare
Smart energy
grids
Smart retail
1 Trillion
25 Billion
10x
70¢ per $1
Devices will be
connected to
the internet by
2011
Global trading
systems are under
extreme stress,
handling billions of
market data
messages each day
Digital data is
projected to grow
tenfold from 2007
to 2011.
70% on average is
spent on
maintaining current
IT infrastructure
versus adding new
capabilities
•
IT infrastructure must grow to meet these demands
global scope, processing scale, efficiency
3
Hardware Trends
 Increasing transistor density
 Clock Speed leveling off
 More number of cores
 Non-Uniform Memory Access
 Main memory getting larger
4
In 2010 POWER Systems Brings Massive Parallelism
POWER7™
4 threads/core
8 cores/chip
32 sockets/server
Threads
1024 threads
POWER5™
POWER4™
1 thread/core
2 cores/chip
16 sockets/server
2 threads/core
2 cores/chip
32 sockets/server
128 threads
POWER6™
2 threads/core
2 cores/chip
32 sockets/server
128 threads
32 threads
2001
180 nm
2004
130 nm
2007
65 nm
2010
45 nm
5
Agenda
• Evolution of Processor Architecture
• Why should I care?
• Think about Scalability
• How to exploit
• Parallelism in Java
• JVM optimizations for multi-core scalability
6
Why should I care?
Your
application may be re-used
Better
performance
 Better
leverage additional resources
 Cores,
hardware threads, memory etc
7
Think about scalability
• Serial bottlenecks inhibit scalability
• Organize your application into parallel tasks
– Consider TaskExecutor API
– Too many threads can be just as bad as too few
• Do not rely on JVM to discover opportunities
– No automatic parallelization
– Java class libraries do not exploit vector processor
capabilities
8
Think about scalability
• Load imbalance
– Workload not evenly distributed
– Consider breaking large tasks into smaller ones
– Change serial algorithms to parallel ones
• Tracing and I/O
– Bottleneck unless infrequent updates or log is
striped (RAID)
– Blocking disk/console I/O inhibit scalability
9
Synchronization and locking
• J9's Three-tiered locking
– Spin
– Yield
– OS
• Avoid synchronization in static methods
• Consider breaking long synchronized blocks into
several smaller ones
– May be bad if results in many context switches
• Java Lock Monitor (JLM) tool can help
http://perfinsp.sourceforge.net/jlm.html
10
Synchronization and locking
• Volatiles
– Compiler will not cache the value
– Creates memory barrier
• Avoid synchronized container classes
– Building scalable data structures is difficult
– Use java.util.concurrent (j/u/c)
• Non-blocking object access
– Possible with j/u/c
11
Agenda
• Evolution of Processor Architecture
• Why should I care?
• Think about Scalability
• How to exploit
• Parallelism in Java
• JVM optimizations for multi-core scalability
12
java.util.concurrent package
• Introduced in Java SE 5
• Alternative strong synchronization
• Lighter weight, better scalability
– Comparing to intrinsic locks
• java.util.concurrent.atomic.*
• java.util.concurrent.locks.*
• ConcurrentCollections
• Synchronizers
• TaskExecutor
13
j/u/c/atomic.*
• Atomic primitives
• Strong form of synchronization
– But does not use lock – non blocking
– Exploit atomic instructions such as compareand-swap in hardware
• Supports compounded actions
•
•
•
•
•
•
AtomicBoolean
AtomicInteger
AtomicIntegerArray
AtomicIntegerFieldUpdater
AtomicLong
AtomicLongArray
•
•
•
•
•
•
AtomicLongFieldUpdater
AtomicMarkableReference
AtomicReference
AtomicReferenceArray
AtomicReferenceFieldUpdater
AtomicStampedReference
14
j/u/c/atomic.*
• Getter and setters
– get
– set
– lazySet
• Updates
– getAndSet
– getAndAdd/getAndIncrement/getAndDecrement
– addAndGet/incrementAndGet/decrementAndGet
• CAS
– compareAndSet/weakCompareAndSet
• Conversions
– toString, intValue, longValue, floatValue, doubleValue
15
j/u/c/locks.*
• Problems with intrinsic locks
– Impossible to back off from a lock attempt
• Deadlock
– Lack of features
• Read vs write
• Fairness policies
– Block-structured
• Must lock and release in the same method
• j/u/c/locks
– Greater flexibility for locks and conditions
• Non-block-structured
– Provides reader-writer locks
• Why block other readers?
• Better scalability
16
j/u/c/locks.*
Interfaces:
– Condition
– Lock
– ReadWriteLock
Classes:
– ReentrantLock
– ReentrantReadWriteLock
– LockSupport
– AbstractQueuedSynchronizer
17
j/u/c.* - Concurrent Collections
Concurrent, thread safe implementations of
several collections
HashMap → ConcurrentHashMap
TreeMap → ConcurrentSkipListMap
ArrayList → CopyOnWriteArrayList
ArraySet → CopyOnWriteArraySet
Queues → ConcurrentLinkedQueue or one
of the blocking queues
18
Strains on the VM
• Excessive use of temporary memory can lead to
increased garbage collector activity
– Stop the world GC pauses the application
• Excessive class loading
– Updating class hierarchy
– Invalidating JIT optimizations
– Consider creating a “startup” phase
• Transitions between Java and native code
– VM access lock
19
Memory Footprint
• Little control over object allocation in Java
– Small short lived objects are easier to cache
– Large long lived objects likely to cause cache
misses
– Memory Analysis Tool (MAT) can help
• Consider using large pages for TLB misses
– -Xlp, requires OS support
• Tune your heap settings
– Heap lock contention with flat heap
20
Affinitizing JVMs
• Can exploit cache hierarchy on a subset of
cores
• JVM working set can fit within the physical
memory of a single node in a NUMA system
• Linux: taskset, numactl
• Windows: start
21
Is my application scalable?
• Low CPU means resources are not maximized
– Evaluate if application has too few/many threads
– Locks and synchronization
– Network connections, I/O
– Thrashing
• working set is too large for physical memory
• High CPU is generally good, as long as resources are spent in
application threads, doing meaningful work
• Evaluate where time is being spent
– Garbage collection
– VM/JIT
– OS Kernel functions
– Other processes
• Tune, tune, tune
22
Write Once, Tune Everywhere
• HealthCenter, GCMV, MAT
http://www.ibm.com/developerworks/java/jdk/tools/
• Dependence on operating System
– Memory allocation
– Socket layer
• Tune for hardware capabilities
– How many cores? How much memory?
– What is the limit on network access?
– Are there storage bottlenecks?
23
Agenda
• Evolution of Processor Architecture
• Why should I care?
• Think about Scalability
• How to exploit
• Parallelism in Java
• JVM optimizations for multi-core scalability
24
IBM Java Execution Model is Built for Parallelism
JIT
Compiler
Garbage
Collector
• Manages memory on behalf of the
application
• Must balance throughput against
observed pauses
• Exploits many multiple hardware
threads
• Generates high performance code
for application threads
• Customizes execution to underlying
hardware
• Optimizes locking performance
• Asynchronous compilation thread
Application
Threads
• Java software threads are
executed on multiple hardware
threads
• Thread safe libraries with scalable
concurrency support for parallel
programming
25
Configurable Garbage Collection policies
 Multiple policies to match varying user requirements
– Pause time, Throughput, Memory footprint and GC overhead
 All modes exploit parallel execution
– Dynamic adaptation to number of available hardware cores &
threads
– GC scalability independent from user application scalability
– Very low overhead (<3%) on typical workloads
26
How do GC policies compare? - optthruput
Optimize Throughput
•Highly parallel GC + streamlined application thread execution
•May cause longer pause times
Java
GC
-Xgcpolicy:optthruput
Thread 1
Thread 2
Thread 3
Thread n
Time
Picture is only illustrative and doesn’t reflect any particular real-life application. The purpose is to
show theoretical differences in pause times between GC policies.
27
How do GC policies compare? - optavgpause
Optimize Pause Time
•GC cleans up concurrently with application thread execution
•Sacrifice some throughput to reduce average pause times
Java
GC
Concurrent Tracing
-Xgcpolicy:optavgpause
Thread 1
Thread 2
Thread 3
Thread n
Time
Picture is only illustrative and doesn’t reflect any particular real-life application. The purpose is to
show theoretical differences in pause times between GC policies.
28
How do GC policies compare? - gencon
Balanced
•Clean up many short-lived objects concurrent with application threads
•Some pauses needed to collect longer-lived objects
Java
Global GC
Scavenge GC
Concurrent Tracing
-Xgcpolicy:gencon
Thread 1
Thread 2
Thread 3
Thread n
Time
Picture is only illustrative and doesn’t reflect any particular real-life application. The purpose is to
show theoretical differences in pause times between GC policies.
29
How do GC policies compare? - subpools
Scalable
•Scalable GC focused on the larger multiprocessor machines
•Improved object allocation algorithm
•May not be appropriate for small-to-midsize configurations
–Xgcpolicy:subpool
• Uses multiple free lists
•Tries to predict the size of future allocation requests based
on earlier allocation requests.
•Recreates free lists at the end of each GC based on these
predictions.
•While allocating objects on the heap, free chunks are
chosen using a “best fit” method, as against the “first fit”
method used in other algorithms.
•Concurrent marking is disabled
30
JVM optimizations for multi-core scalability
 Lock removal across JVM and class libraries
 java.util.concurrent package optimizations
 Better working set for cache efficiency
 Stack allocation
 Remove/optimize synchronization
 Thread local storage for send/receive buffers
 Non-blocking containers
 Asynch JIT compilation on a separate thread
 Right-sized application runtimes
31
Russian
Gracias
Spanish
Obrigado
Brazilian Portuguese
Merci
French
Thank You
Arabic
Traditional Chinese
Simplified Chinese
Thai
Questions?
Korean
Grazie
Italian
Email: [email protected]
Danke
Japanese
http://www.ibm.com/developerworks/java/
German
32
Special notices
© IBM Corporation 2010. All Rights Reserved.
The workshops, sessions and materials have been prepared by IBM or the session speakers and reflect their own views. They are provided for informational purposes only, and
are neither intended to, nor shall have the effect of being, legal or other guidance or advice to any participant. While efforts were made to verify the completeness and
accuracy of the information contained in this presentation, it is provided AS IS without warranty of any kind, express or implied. IBM shall not be responsible for any
damages arising out of the use of, or otherwise related to, this presentation or any other materials. Nothing contained in this presentation is intended to, nor shall have the
effect of, creating any warranties or representations from IBM or its suppliers or licensors, or altering the terms and conditions of the applicable license agreement governing
the use of IBM software.
References in this presentation to IBM products, programs, or services do not imply that they will be available in all countries in which IBM operates. Product release dates and/or
capabilities referenced in this presentation may change at any time at IBM’s sole discretion based on market opportunities or other factors, and are not intended to be a
commitment to future product or feature availability in any way. Nothing contained in these materials is intended to, nor shall have the effect of, stating or implying that any
activities undertaken by you will result in any specific sales, revenue growth or other results.
Performance is based on measurements and projections using standard IBM benchmarks in a controlled environment. The actual throughput or performance that any user will
experience will vary depending upon many factors, including considerations such as the amount of multiprogramming in the user's job stream, the I/O configuration, the
storage configuration, and the workload processed. Therefore, no assurance can be given that an individual user will achieve results similar to those stated here.
All customer examples described are presented as illustrations of how those customers have used IBM products and the results they may have achieved. Actual environmental
costs and performance characteristics may vary by customer.
The following are trademarks of the International Business Machines Corporation in the United States and/or other countries:
ibm.com/legal/copytrade.shtmlAIX, CICS, CICSPlex, DataPower, DB2, DB2 Universal Database, i5/OS, IBM, the IBM logo, IMS/ESA, Power Systems, Lotus, OMEGAMON,
OS/390, Parallel Sysplex, pureXML, Rational, Redbooks, Sametime, SMART SOA, System z , Tivoli, WebSphere, and z/OS.
A current list of IBM trademarks is available on the Web at “Copyright and trademark information” at ibm.com/legal/copytrade.shtml.
Adobe, the Adobe logo, PostScript, and the PostScript logo are either registered trademarks or trademarks of Adobe Systems Incorporated in the United States, and/or other
countries.
IT Infrastructure Library is a registered trademark of the Central Computer and Telecommunications Agency which is now part of the Office of Government Commerce
Java and all Java-based trademarks are trademarks of Sun Microsystems, Inc. in the United States, other countries, or both.
Microsoft and Windows are trademarks of Microsoft Corporation in the United States, other countries, or both.
ITIL is a registered trademark, and a registered community trademark of the Office of Government Commerce, and is registered in the U.S. Patent and Trademark Office
Intel and Pentium are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries.
UNIX is a registered trademark of The Open Group in the United States and other countries.
Linux is a registered trademark of Linus Torvalds in the United States, other countries, or both.
33