Apache logging services logo Apache log4j logo

Asynchronous Loggers for Low-Latency Logging

Asynchronous logging can improve your application's performance by executing the I/O operations in a separate thread. Log4j 2 makes a number of improvements in this area.

  • Asynchronous Loggers are a new addition in Log4j 2. Their aim is to return from the call to Logger.log to the application as soon as possible. You can choose between making all Loggers asynchronous or using a mixture of synchronous and asynchronous Loggers. Making all Loggers asynchronous will give the best performance, while mixing gives you more flexibility.
  • LMAX Disruptor technology. Asynchronous Loggers internally use the Disruptor, a lock-free inter-thread communication library, instead of queues, resulting in higher throughput and lower latency.
  • As part of the work for Async Loggers, Asynchronous Appenders have been enhanced to flush to disk at the end of a batch (when the queue is empty). This produces the same result as configuring "immediateFlush=true", that is, all received log events are always available on disk, but is more efficient because it does not need to touch the disk on each and every log event. (Async Appenders use ArrayBlockingQueue internally and do not need the disruptor jar on the classpath.)

Trade-offs

Although asynchronous logging can give significant performance benefits, there are situations where you may want to choose synchronous logging. This section describes some of the trade-offs of asynchronous logging.

Benefits

  • Higher peak throughput. With an asynchronous logger your application can log messages at 6 - 68 times the rate of a synchronous logger.

    This is especially interesting for applications that occasionally need to log bursts of messages. Async logging can help prevent or dampen latency spikes by shortening the wait time until the next message can be logged. If the queue size is configured large enough to handle the burst, asynchronous logging will help prevent your application from falling behind (as much) during a sudden increase of activity.

  • Lower logging response time latency. Response time latency is the time it takes for a call to Logger.log to return under a given workload. Asynchronous Loggers have consistently lower latency than synchronous loggers or even queue-based asynchronous appenders.
Drawbacks
  • Error handling. If a problem happens during the logging process and an exception is thrown, it is less easy for an asynchronous logger or appender to signal this problem to the application. This can partly be alleviated by configuring an ExceptionHandler, but this may still not cover all cases. For this reason, if logging is part of your business logic, for example if you are using Log4j as an audit logging framework, we would recommend to synchronously log those audit messages. (Note that you can still combine them and use asynchronous logging for debug/trace logging in addition to synchronous logging for the audit trail.)
  • In some rare cases, care must be taken with mutable messages. Most of the time you don't need to worry about this. Log4 will ensure that log messages like logger.debug("My object is {}", myObject) will use the state of the myObject parameter at the time of the call to logger.debug(). The log message will not change even if myObject is modified later. It is safe to asynchronously log mutable objects because most Message implementations built-in to Log4j take a snapshot of the parameters. There are some exceptions however: MapMessage and StructuredDataMessage are mutable by design: fields can be added to these messages after the message object was created. These messages should not be modified after they are logged with asynchronous loggers or asynchronous appenders; you may or may not see the modifications in the resulting log output. Similarly, custom Message implementations should be designed with asynchronous use in mind, and either take a snapshot of their parameters at construction time, or document their thread-safety characteristics.
  • If your application is running in an environment where CPU resources are scarce, like a machine with one CPU with a single core, starting another thread is not likely to give better performance.
  • If the sustained rate at which your application is logging messages is faster than the maximum sustained throughput of the underlying appender, the queue will fill up and the application will end up logging at the speed of the slowest appender. If this happens, consider selecting a faster appender, or logging less. If neither of these is an option, you may get better throughput and fewer latency spikes by logging synchronously.

Making All Loggers Asynchronous

Log4j-2.9 and higher require disruptor-3.3.4.jar or higher on the classpath. Prior to Log4j-2.9, disruptor-3.0.0.jar or higher was required.

This is simplest to configure and gives the best performance. To make all loggers asynchronous, add the disruptor jar to the classpath and set the system property Log4jContextSelector to org.apache.logging.log4j.core.async.AsyncLoggerContextSelector.

By default, location is not passed to the I/O thread by asynchronous loggers. If one of your layouts or custom filters needs location information, you need to set "includeLocation=true" in the configuration of all relevant loggers, including the root logger.

A configuration that does not require location might look like:

<?xml version="1.0" encoding="UTF-8"?>

<!-- Don't forget to set system property
-DLog4jContextSelector=org.apache.logging.log4j.core.async.AsyncLoggerContextSelector
     to make all loggers asynchronous. -->

<Configuration status="WARN">
  <Appenders>
    <!-- Async Loggers will auto-flush in batches, so switch off immediateFlush. -->
    <RandomAccessFile name="RandomAccessFile" fileName="async.log" immediateFlush="false" append="false">
      <PatternLayout>
        <Pattern>%d %p %c{1.} [%t] %m %ex%n</Pattern>
      </PatternLayout>
    </RandomAccessFile>
  </Appenders>
  <Loggers>
    <Root level="info" includeLocation="false">
      <AppenderRef ref="RandomAccessFile"/>
    </Root>
  </Loggers>
</Configuration>

When AsyncLoggerContextSelector is used to make all loggers asynchronous, make sure to use normal <root> and <logger> elements in the configuration. The AsyncLoggerContextSelector will ensure that all loggers are asynchronous, using a mechanism that is different from what happens when you configure <asyncRoot> or <asyncLogger>. The latter elements are intended for mixing async with sync loggers. If you use both mechanisms together you will end up with two background threads, where your application passes the log message to thread A, which passes the message to thread B, which then finally logs the message to disk. This works, but there will be an unnecessary step in the middle.

There are a few system properties you can use to control aspects of the asynchronous logging subsystem. Some of these can be used to tune logging performance.

The below properties can also be specified by creating a file named log4j2.component.properties and including this file in the classpath of the application.

System Properties to configure all asynchronous loggers
System Property Default Value Description
AsyncLogger.ExceptionHandler default handler Fully qualified name of a class that implements the com.lmax.disruptor.ExceptionHandler interface. The class needs to have a public zero-argument constructor. If specified, this class will be notified when an exception occurs while logging the messages.

If not specified, the default exception handler will print a message and stack trace to the standard error output stream.

AsyncLogger.RingBufferSize 256 * 1024 Size (number of slots) in the RingBuffer used by the asynchronous logging subsystem. Make this value large enough to deal with bursts of activity. The minimum size is 128. The RingBuffer will be pre-allocated at first use and will never grow or shrink during the life of the system.

When the application is logging faster than the underlying appender can keep up with for a long enough time to fill up the queue, the behavious is determined by the AsyncQueueFullPolicy.

AsyncLogger.WaitStrategy Timeout Valid values: Block, Timeout, Sleep, Yield.
Block is a strategy that uses a lock and condition variable for the I/O thread waiting for log events. Block can be used when throughput and low-latency are not as important as CPU resource. Recommended for resource constrained/virtualised environments.
Timeout is a variation of the Block strategy that will periodically wake up from the lock condition await() call. This ensures that if a notification is missed somehow the consumer thread is not stuck but will recover with a small latency delay (default 10ms).
Sleep is a strategy that initially spins, then uses a Thread.yield(), and eventually parks for the minimum number of nanos the OS and JVM will allow while the I/O thread is waiting for log events. Sleep is a good compromise between performance and CPU resource. This strategy has very low impact on the application thread, in exchange for some additional latency for actually getting the message logged.
Yield is a strategy that uses a Thread.yield() for waiting for log events after an initially spinning. Yield is a good compromise between performance and CPU resource, but may use more CPU than Sleep in order to get the message logged to disk sooner.
AsyncLogger.ThreadNameStrategy CACHED Valid values: CACHED, UNCACHED.
By default, AsyncLogger caches the thread name in a ThreadLocal variable to improve performance. Specify the UNCACHED option if your application modifies the thread name at runtime (with Thread.currentThread().setName()) and you want to see the new thread name reflected in the log.
log4j.Clock SystemClock

Implementation of the org.apache.logging.log4j.core.util.Clock interface that is used for timestamping the log events when all loggers are asynchronous.
By default, System.currentTimeMillis is called on every log event.

CachedClock is an optimization intended for low-latency applications where time stamps are generated from a clock that updates its internal time in a background thread once every millisecond, or every 1024 log events, whichever comes first. This reduces logging latency a little, at the cost of some precision in the logged time stamps. Unless you are logging many events, you may see "jumps" of 10-16 milliseconds between log time stamps. WEB APPLICATION WARNING: The use of a background thread may cause issues for web applications and OSGi applications so CachedClock is not recommended for this kind of applications.

You can also specify the fully qualified class name of a custom class that implements the Clock interface.

There are also a few system properties that can be used to maintain application throughput even when the underlying appender cannot keep up with the logging rate and the queue is filling up. See the details for system properties log4j2.AsyncQueueFullPolicy and log4j2.DiscardThreshold.

Mixing Synchronous and Asynchronous Loggers

Log4j-2.9 and higher require disruptor-3.3.4.jar or higher on the classpath. Prior to Log4j-2.9, disruptor-3.0.0.jar or higher was required. There is no need to set system property "Log4jContextSelector" to any value.

Synchronous and asynchronous loggers can be combined in configuration. This gives you more flexibility at the cost of a slight loss in performance (compared to making all loggers asynchronous). Use the <asyncRoot> or <asyncLogger> configuration elements to specify the loggers that need to be asynchronous. A configuration can contain only one root logger (either a <root> or an <asyncRoot> element), but otherwise async and non-async loggers may be combined. For example, a configuration file containing <asyncLogger> elements can also contain <root> and <logger> elements for the synchronous loggers.

By default, location is not passed to the I/O thread by asynchronous loggers. If one of your layouts or custom filters needs location information, you need to set "includeLocation=true" in the configuration of all relevant loggers, including the root logger.

A configuration that mixes asynchronous loggers might look like:

<?xml version="1.0" encoding="UTF-8"?>

<!-- No need to set system property "Log4jContextSelector" to any value
     when using <asyncLogger> or <asyncRoot>. -->

<Configuration status="WARN">
  <Appenders>
    <!-- Async Loggers will auto-flush in batches, so switch off immediateFlush. -->
    <RandomAccessFile name="RandomAccessFile" fileName="asyncWithLocation.log"
              immediateFlush="false" append="false">
      <PatternLayout>
        <Pattern>%d %p %class{1.} [%t] %location %m %ex%n</Pattern>
      </PatternLayout>
    </RandomAccessFile>
  </Appenders>
  <Loggers>
    <!-- pattern layout actually uses location, so we need to include it -->
    <AsyncLogger name="com.foo.Bar" level="trace" includeLocation="true">
      <AppenderRef ref="RandomAccessFile"/>
    </AsyncLogger>
    <Root level="info" includeLocation="true">
      <AppenderRef ref="RandomAccessFile"/>
    </Root>
  </Loggers>
</Configuration>

There are a few system properties you can use to control aspects of the asynchronous logging subsystem. Some of these can be used to tune logging performance.

The below properties can also be specified by creating a file named log4j2.component.properties and including this file in the classpath of the application.

System Properties to configure mixed asynchronous and normal loggers
System Property Default Value Description
AsyncLoggerConfig.ExceptionHandler default handler Fully qualified name of a class that implements the com.lmax.disruptor.ExceptionHandler interface. The class needs to have a public zero-argument constructor. If specified, this class will be notified when an exception occurs while logging the messages.

If not specified, the default exception handler will print a message and stack trace to the standard error output stream.

AsyncLoggerConfig.RingBufferSize 256 * 1024 Size (number of slots) in the RingBuffer used by the asynchronous logging subsystem. Make this value large enough to deal with bursts of activity. The minimum size is 128. The RingBuffer will be pre-allocated at first use and will never grow or shrink during the life of the system.

When the application is logging faster than the underlying appender can keep up with for a long enough time to fill up the queue, the behavious is determined by the AsyncQueueFullPolicy.

AsyncLoggerConfig.WaitStrategy Timeout Valid values: Block, Timeout, Sleep, Yield.
Block is a strategy that uses a lock and condition variable for the I/O thread waiting for log events. Block can be used when throughput and low-latency are not as important as CPU resource. Recommended for resource constrained/virtualised environments.
Timeout is a variation of the Block strategy that will periodically wake up from the lock condition await() call. This ensures that if a notification is missed somehow the consumer thread is not stuck but will recover with a small latency delay (default 10ms).
Sleep is a strategy that initially spins, then uses a Thread.yield(), and eventually parks for the minimum number of nanos the OS and JVM will allow while the I/O thread is waiting for log events. Sleep is a good compromise between performance and CPU resource. This strategy has very low impact on the application thread, in exchange for some additional latency for actually getting the message logged.
Yield is a strategy that uses a Thread.yield() for waiting for log events after an initially spinning. Yield is a good compromise between performance and CPU resource, but may use more CPU than Sleep in order to get the message logged to disk sooner.

There are also a few system properties that can be used to maintain application throughput even when the underlying appender cannot keep up with the logging rate and the queue is filling up. See the details for system properties log4j2.AsyncQueueFullPolicy and log4j2.DiscardThreshold.

Location, location, location...

If one of the layouts is configured with a location-related attribute like HTML locationInfo, or one of the patterns %C or $class, %F or %file, %l or %location, %L or %line, %M or %method, Log4j will take a snapshot of the stack, and walk the stack trace to find the location information.

This is an expensive operation: 1.3 - 5 times slower for synchronous loggers. Synchronous loggers wait as long as possible before they take this stack snapshot. If no location is required, the snapshot will never be taken.

However, asynchronous loggers need to make this decision before passing the log message to another thread; the location information will be lost after that point. The performance impact of taking a stack trace snapshot is even higher for asynchronous loggers: logging with location is 30-100 times slower than without location. For this reason, asynchronous loggers and asynchronous appenders do not include location information by default.

You can override the default behaviour in your logger or asynchronous appender configuration by specifying includeLocation="true".

Asynchronous Logging Performance

The throughput performance results below were derived from running the PerfTest, MTPerfTest and PerfTestDriver classes which can be found in the Log4j 2 unit test source directory. For throughput tests, the methodology used was:

  • First, warm up the JVM by logging 200,000 log messages of 500 characters.
  • Repeat the warm-up 10 times, then wait 10 seconds for the I/O thread to catch up and buffers to drain.
  • Measure how long it takes to execute 256 * 1024 / threadCount calls to Logger.log and express the result in messages per second.
  • Repeat the test 5 times and average the results.

The results below were obtained with log4j-2.0-beta5, disruptor-3.0.0.beta3, log4j-1.2.17 and logback-1.0.10.

Logging Peak Throughput

The graph below compares the throughput of synchronous loggers, asynchronous appenders and asynchronous loggers. This is the total throughput of all threads together. In the test with 64 threads, asynchronous loggers are 12 times faster than asynchronous appenders, and 68 times faster than synchronous loggers.

Asynchronous loggers' throughput increases with the number of threads, whereas both synchronous loggers and asynchronous appenders have more or less constant throughput regardless of the number of threads that are doing the logging.

Async loggers have much higher throughput than sync loggers.

Asynchronous Throughput Comparison with Other Logging Packages

We also compared peak throughput of asynchronous loggers to the synchronous loggers and asynchronous appenders available in other logging packages, specifically log4j-1.2.17 and logback-1.0.10, with similar results. For asynchronous appenders, total logging throughput of all threads together remains roughly constant when adding more threads. Asynchronous loggers make more effective use of the multiple cores available on the machine in multi-threaded scenarios.

Async loggers have the highest throughput.

On Solaris 10 (64bit) with JDK1.7.0_06, 4-core Xeon X5570 dual CPU @2.93Ghz with hyperthreading switched on (16 virtual cores):

Throughput per thread in messages/second
Logger 1 thread 2 threads 4 threads 8 threads 16 threads 32 threads 64 threads
Log4j 2: Loggers all asynchronous 2,652,412 909,119 776,993 516,365 239,246 253,791 288,997
Log4j 2: Loggers mixed sync/async 2,454,358 839,394 854,578 597,913 261,003 216,863 218,937
Log4j 2: Async Appender 1,713,429 603,019 331,506 149,408 86,107 45,529 23,980
Log4j1: Async Appender 2,239,664 494,470 221,402 109,314 60,580 31,706 14,072
Logback: Async Appender 2,206,907 624,082 307,500 160,096 85,701 43,422 21,303
Log4j 2: Synchronous 273,536 136,523 67,609 34,404 15,373 7,903 4,253
Log4j1: Synchronous 326,894 105,591 57,036 30,511 13,900 7,094 3,509
Logback: Synchronous 178,063 65,000 34,372 16,903 8,334 3,985 1,967

On Windows 7 (64bit) with JDK1.7.0_11, 2-core Intel i5-3317u CPU @1.70Ghz with hyperthreading switched on (4 virtual cores):

Throughput per thread in messages/second
Logger 1 thread 2 threads 4 threads 8 threads 16 threads 32 threads
Log4j 2: Loggers all asynchronous 1,715,344 928,951 1,045,265 1,509,109 1,708,989 773,565
Log4j 2: Loggers mixed sync/async 571,099 1,204,774 1,632,204 1,368,041 462,093 908,529
Log4j 2: Async Appender 1,236,548 1,006,287 511,571 302,230 160,094 60,152
Log4j1: Async Appender 1,373,195 911,657 636,899 406,405 202,777 162,964
Logback: Async Appender 1,979,515 783,722 582,935 289,905 172,463 133,435
Log4j 2: Synchronous 281,250 225,731 129,015 66,590 34,401 17,347
Log4j1: Synchronous 147,824 72,383 32,865 18,025 8,937 4,440
Logback: Synchronous 149,811 66,301 32,341 16,962 8,431 3,610

Response Time Latency

This section has been rewritten with the Log4j 2.6 release. The previous version only reported service time instead of response time. See the response time side bar on the performance page on why this is too optimistic. Furthermore the previous version reported average latency, which does not make sense since latency is not a normal distribution. Finally, the previous version of this section only reported the maximum latency of up to 99.99% of the measurements, which does not tell you how bad the worst 0.01% were. This is unfortunate because often the "outliers" are all that matter when it comes to response time. From this release we will try to do better and report response time latency across the full range of percentages, including all the outliers. Our thanks to Gil Tene for his How NOT to measure latency presentation. (Now we know why this is also known as the "Oh s#@t!" presentation.)

Response time is how long it takes to log a message under a certain load. What is often reported as latency is actually service time: how long it took to perform the operation. This hides the fact that a single spike in service time adds queueing delay for many of the subsequent operations. Service time is easy to measure (and often looks good on paper) but is irrelevant for users since it omits the time spent waiting for service. For this reason we report response time: service time plus wait time.

The response time test results below were all derived from running the ResponseTimeTest class which can be found in the Log4j 2 unit test source directory. If you want to run these tests yourself, here are the command line options we used:

  • -Xms1G -Xmx1G (prevent heap resizing during the test)
  • -DLog4jContextSelector=org.apache.logging.log4j.core.async.AsyncLoggerContextSelector -DAsyncLogger.WaitStrategy=busyspin (to use Async Loggers. The BusySpin wait strategy reduces some jitter.)
  • classic mode: -Dlog4j2.enable.threadlocals=false -Dlog4j2.enable.direct.encoders=false
    garbage-free mode: -Dlog4j2.enable.threadlocals=true -Dlog4j2.enable.direct.encoders=true
  • -XX:CompileCommand=dontinline,org.apache.logging.log4j.core.async.perftest.NoOpIdleStrategy::idle
  • -verbose:gc -XX:+PrintGCDetails -XX:+PrintGCDateStamps -XX:+PrintTenuringDistribution -XX:+PrintGCApplicationConcurrentTime -XX:+PrintGCApplicationStoppedTime (to eyeball GC and safepoint pauses)

The graph below compares response time latency of the ArrayBlockingQueue-based asynchronous appenders in Logback 1.1.7, Log4j 1.2.17 to the various options for asynchronous logging that Log4j 2.6 offers. Under a workload of 128,000 messages per second, using 16 threads (each logging at a rate of 8,000 messages per second), we see that Logback 1.1.7, Log4j 1.2.17 experience latency spikes that are orders of magnitude larger than Log4j 2.

When 16 threads generate a total workload of 128,000 msg/sec, Logback 1.1.7 and                Log4j 1.2.17 experience latency spikes that are orders of magnitude larger than Log4j 2

The graph below zooms in on the Log4j 2 results for the same test. We see that the worst-case response time is highest for the ArrayBlockingQueue-based Async Appender. Garbage-free async loggers have the best response time behaviour.

Under The Hood

Asynchronous Loggers are implemented using the LMAX Disruptor inter-thread messaging library. From the LMAX web site:

... using queues to pass data between stages of the system was introducing latency, so we focused on optimising this area. The Disruptor is the result of our research and testing. We found that cache misses at the CPU-level, and locks requiring kernel arbitration are both extremely costly, so we created a framework which has "mechanical sympathy" for the hardware it's running on, and that's lock-free.

LMAX Disruptor internal performance comparisons with java.util.concurrent.ArrayBlockingQueue can be found here.