IntelliJ IDEA Architecture and Performance Dmit ry Jemerov Development Lead JetBrains, Inc. ht tp://www.jetbrains.com/

Learn about performance optimization techniques used in IntelliJ IDEA Understand the factors affecting IntelliJ IDEA performance in normal use

Learn about performance optimization techniques used in IntelliJ IDEA

Understand the factors affecting IntelliJ IDEA performance in normal use

Agenda Elements of Architecture Performance Tips

Elements of Architecture

Performance Tips

Component Model Hierarchy of PicoContainer containers 3 levels of components: application, project and module Components have strictly defined lifecycle Registration through XML configuration files (typically)

Hierarchy of PicoContainer containers

3 levels of components: application, project and module

Components have strictly defined lifecycle

Registration through XML configuration files (typically)

Services: Lazy Loading No strict lifecycle: loaded on demand Key for loading service is FQN of service interface Not even classloading happens before service is accessed Service may have persistent settings which are automatically loaded on instantiation

No strict lifecycle: loaded on demand

Key for loading service is FQN of service interface

Not even classloading happens before service is accessed

Service may have persistent settings which are automatically loaded on instantiation

Disposer Hierarchy Tree of dependencies between objects Disposing of parent object also disposes of all its children Relationships stored outside of object: no common base class required Typical actions on dispose: Detaching listeners Stopping timers/background threads UI component cleanup

Tree of dependencies between objects

Disposing of parent object also disposes of all its children

Relationships stored outside of object: no common base class required

Typical actions on dispose:

Detaching listeners

Stopping timers/background threads

UI component cleanup

Persistent File System Complete snapshot of files under project stored in single file Improves startup performance Directory tree walk is very expensive, especially on Windows Provides old content when file is changed Required for update of indexes Supports custom persistent attributes

Complete snapshot of files under project stored in single file

Improves startup performance

Directory tree walk is very expensive, especially on Windows

Provides old content when file is changed

Required for update of indexes

Supports custom persistent attributes

File System Synchronization On startup, background synchronization of all files under project is performed Watching for file system notifications ReadDirectoryChangesW() on Windows /dev/fsevents on Mac (fslogger) Linux has no API for recursive watching, requires keeping a file handle for every directory under project

On startup, background synchronization of all files under project is performed

Watching for file system notifications

ReadDirectoryChangesW() on Windows

/dev/fsevents on Mac (fslogger)

Linux has no API for recursive watching, requires keeping a file handle for every directory under project

Indexing Framework Indexer builds list of key/value pairs from file content Keys and values are arbitrary beans Indexer provides binary serialization for keys and values Application-level same file under different projects indexed only once

Indexer builds list of key/value pairs from file content

Keys and values are arbitrary beans

Indexer provides binary serialization for keys and values

Application-level

same file under different projects indexed only once

Indexing Framework Usages Word index Key: hashcode of word Value: occurrence mask identifier, comment, string literal Used for Find in Path, Find Usages etc. TODO index File name index etc.

Word index

Key: hashcode of word

Value: occurrence mask

identifier, comment, string literal

Used for Find in Path, Find Usages etc.

TODO index

File name index

etc.

PSI Program Source Interface Read/write access to program syntax tree From directories down to individual tokens Works across languages Some elements (whitespace, comment etc.) used by all languages

Program Source Interface

Read/write access to program syntax tree

From directories down to individual tokens

Works across languages

Some elements (whitespace, comment etc.) used by all languages

Building PSI from Source Lexer produces stream of tokens Lexers usually generated with Flex Parser builds syntax tree Usually hand-written recursive descent Table-driven parsers seldom provide adequate error recovery Language creates PSI elements for syntax tree nodes

Lexer produces stream of tokens

Lexers usually generated with Flex

Parser builds syntax tree

Usually hand-written recursive descent

Table-driven parsers seldom provide adequate error recovery

Language creates PSI elements for syntax tree nodes

Updating PSI Syntax highlighting based on lexer Synchronous and incremental PSI updates triggered by background analysis Starts after 300 ms of inactivity Cancelled by next user action Syntax tree (AST) is built for entire file PSI is replaced for changed parts of AST

Syntax highlighting based on lexer

Synchronous and incremental

PSI updates triggered by background analysis

Starts after 300 ms of inactivity

Cancelled by next user action

Syntax tree (AST) is built for entire file

PSI is replaced for changed parts of AST

PSI Stubs Compact binary serialization of externally visible interface of file Can be provided for any language Must be sufficient for resolving imported declarations without parsing imported files PSI elements can exist in two modes: stub-based and AST-based Switching from stubs to AST is transparent for the client

Compact binary serialization of externally visible interface of file

Can be provided for any language

Must be sufficient for resolving imported declarations without parsing imported files

PSI elements can exist in two modes: stub-based and AST-based

Switching from stubs to AST is transparent for the client

Stub Indexes For a stub element, a number of keys for different indexes can be provided PSI elements can then be retrieved by such keys Sample indexes: Classes by name (Goto Class) Symbols by name (Goto Symbol) Classes by superclass name (Goto Implementations)

For a stub element, a number of keys for different indexes can be provided

PSI elements can then be retrieved by such keys

Sample indexes:

Classes by name (Goto Class)

Symbols by name (Goto Symbol)

Classes by superclass name (Goto Implementations)

Background Analysis Multiple passes with different priorities Faster analysis performed first Separate high-priority passes for highlighting only visible area Minimum affected scope is re-analyzed after change Cancellation on user activity Cancellation flag checked in many low-level operations, exception thrown and caught on top level

Multiple passes with different priorities

Faster analysis performed first

Separate high-priority passes for highlighting only visible area

Minimum affected scope is re-analyzed after change

Cancellation on user activity

Cancellation flag checked in many low-level operations, exception thrown and caught on top level

Multi-Core Support Inspections: Different inspections for same file run in parallel on multiple cores Find Usages: Different occurrences of target identifier are checked in parallel JobScheduler API Starts # of threads depending on # of available CPU cores Distributes work units between threads Waits until all work units processed

Inspections: Different inspections for same file run in parallel on multiple cores

Find Usages: Different occurrences of target identifier are checked in parallel

JobScheduler API

Starts # of threads depending on # of available CPU cores

Distributes work units between threads

Waits until all work units processed

Multi-Core Support Challenges Deadlocks Lock ordering policies must be followed consistently Lock contention Lock granularity must be reduced to ensure that all threads can run in parallel YourKit used for profiling Creating too many sync objects Lock pooling: fixed size pool of locks evenly distributed between all objects

Deadlocks

Lock ordering policies must be followed consistently

Lock contention

Lock granularity must be reduced to ensure that all threads can run in parallel

YourKit used for profiling

Creating too many sync objects

Lock pooling: fixed size pool of locks evenly distributed between all objects

Agenda Elements of Architecture Performance Tips

Elements of Architecture

Performance Tips

Performance Tips: System IDEA is heavily IO-bound Do not store sources, caches or system directories on network drives Disable antivirus scanning for system directory Regularly run disk defragmenter Turn off Windows System Restore

IDEA is heavily IO-bound

Do not store sources, caches or system directories on network drives

Disable antivirus scanning for system directory

Regularly run disk defragmenter

Turn off Windows System Restore

Performance Tips: JVM Use JDK 6 In –server mode, increasing –Xmx helps IDEA’s caches live longer Parallel GC helps on multicore machines However : May cause stability issues. YMMV.

Use JDK 6

In –server mode, increasing –Xmx helps IDEA’s caches live longer

Parallel GC helps on multicore machines

However : May cause stability issues. YMMV.

Performance Tips: IDE Use latest version Disable unused plugins Exclude unneeded folders Set version control to None for generated code Avoid large number of unversioned files

Use latest version

Disable unused plugins

Exclude unneeded folders

Set version control to None for generated code

Avoid large number of unversioned files

Performance Tips: IDE (2) Do not disable “Synchronize files on frame activation” Switch Project view to “Project” rather than “Packages” Disable expensive inspections Disable framework-specific validation on Make Do not use method breakpoints

Do not disable “Synchronize files on frame activation”

Switch Project view to “Project” rather than “Packages”

Disable expensive inspections

Disable framework-specific validation on Make

Do not use method breakpoints

Reporting Performance Issues Enable built-in YourKit profiling -agentlib:yjpagent in idea.exe.vmoptions On by default in EAP builds Capture snapshot Create JIRA issue and attach snapshot http://www.jetbrains.net/jira Also possible to use standalone YourKit For example, for startup profiling

Enable built-in YourKit profiling

-agentlib:yjpagent in idea.exe.vmoptions

On by default in EAP builds

Capture snapshot

Create JIRA issue and attach snapshot

http://www.jetbrains.net/jira

Also possible to use standalone YourKit

For example, for startup profiling

For More Information IntelliJ IDEA: http://www.jetbrains.com/idea/ Plugin Development Kit: http://www.jetbrains.com/idea/plugins/plugin_developers.html Discussion forums: http://www.intellij.net/forums/

IntelliJ IDEA:

http://www.jetbrains.com/idea/

Plugin Development Kit:

http://www.jetbrains.com/idea/plugins/plugin_developers.html

Discussion forums:

http://www.intellij.net/forums/

Q&A Dmitry Jemerov [email_address]

IntelliJ IDEA Architecture and Performance Dmit ry Jemerov Development Lead JetBrains, Inc. ht tp://www.jetbrains.com/