[jira] [Comment Edited] (LOG4J2-1179) Log4j performance documentation

["Remko Popma (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/LOG4J2-1179?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15278315#comment-15278315 ] 

Remko Popma edited comment on LOG4J2-1179 at 5/10/16 3:52 PM:
--------------------------------------------------------------

Similarly, I'm considering postponing the performance comparison of the Log4j 2 Socket Appenders versus the Logback and Log4j1 Socket Appenders unless someone else volunteers to create the benchmarks. 

Note that [the loggly article|https://www.loggly.com/blog/benchmarking-java-logging-frameworks/] that does such a comparison also mentions that events are dropped. This should not happen even for UDP unless you are either sending packets across different data centers or your consumer cannot keep up with the producer. I suspect that the latter was the case with the Loggly benchmark. Does this mean we need to optimize our Tcp/UdpSocketServer implementation?

Anyway, testing socket appender performance under load means not just measuring performance but also tracking if and how many messages are dropped, so this is likely more work than just creating a benchmark.


was (Author: remkop@yahoo.com):
Similarly, I'm considering postponing the performance comparison of the Log4j 2 Socket Appenders versus the Logback and Log4j1 Socket Appenders unless someone else volunteers to create the benchmarks. 

Note that [the loggly article|https://www.loggly.com/blog/benchmarking-java-logging-frameworks/] that does such a comparison also mentions that events are dropped. This should not happen even for UDP unless you are either sending packets across different data centers or your consumer cannot keep up with the producer. I suspect that this was the case with the Loggly benchmark. 

Testing socket appender performance under load means not just measuring performance but also tracking if and how many messages are dropped, so this is more work than just creating a benchmark.

> Log4j performance documentation
> -------------------------------
>
>                 Key: LOG4J2-1179
>                 URL: https://issues.apache.org/jira/browse/LOG4J2-1179
>             Project: Log4j 2
>          Issue Type: Documentation
>          Components: Documentation, Performance Benchmarks
>    Affects Versions: 2.4.1
>            Reporter: Remko Popma
>            Assignee: Remko Popma
>             Fix For: 2.6
>
>         Attachments: ParamMsgThrpt1T.png, ParamMsgThrpt2T.png, ParamMsgThrpt4T.png
>
>
> Reorganize and extend performance data on the site.
> *Async Loggers Manual Page*
> Should be more focussed. Proposed changes:
> (/) Link to Location section in Performance page from Async Loggers page _"Location, location, location..."_ section.
> (/) Similarly, move _"Throughput of Logging With Location (includeLocation="true")"_ table with throughput results to general Performance page. UPDATE: replaced with new data from JMH benchmark.
> (/) Move _"FileAppender vs. RandomAccessFileAppender"_ section to general Performance page. (Again, keep anchors and link to new section on Performance page to avoid breaking links.)
> (/) Rewrite opening paragraph of Async Logger manual page to remove reference to RandomAccessFile appender
> (/) Rewrite section on _Latency_
> * The histogram shows service time (more useful for users is response time: service time + wait time).
> * Bar chart diagram on "average latency" is nonsense. Latency is not a normal distribution so terms like "average latency" don't make sense. Remove this. (A histogram showing the full range of percentiles _does_ make sense.)
> * Bar chart diagram with max of 99.99% of observations is better than average but still has large drawbacks: this is service time (omitting the crucial wait time) and how high are the peaks in the 0.01% we did not report? Better to remove this and instead show a histogram with the full range of percentages.
> *Performance Page*
> (/) Briefly explain about various aspects of "performance": peak measured throughput (what kind of bursts can we deal with?), sustained throughput, and response time (service time + wait time).
> 2. Then show how Log4j 2 compares to the alternatives (Logback, Log4j-1.2 and JUL) on all these three performance dimensions.
> 3. Finally, document some performance trade-offs for Log4j 2 functionality.
> *2. Comparison to alternative logging libraries*
> (/) Peak throughput comparison Async Loggers vs async appenders for bursty logging. 
> (/) Response time comparison of Async Loggers vs async appenders
> (/) Parameterized messages: use these JMH [benchmark results|https://issues.apache.org/jira/browse/LOG4J2-1278?focusedCommentId=15216236&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15216236]? (Looks like parameterized messages are currently quite expensive...)
> (/) compare performance impact of including location between logging libraries
> For various appenders, compare Log4j2 to alternatives with regards to max sustained throughput (and separately, response time).
> (/) [File Appender max sustained thoughput|https://issues.apache.org/jira/browse/LOG4J2-1297?focusedCommentId=15256490&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-15256490]
> (-) File Appender response time comparison
> (?) Socket appender (TCP/UDP)
> (?) Syslog appender (TCP/UDP)
> *3. Log4j 2 functionality performance trade-offs*
> (/) Compare performance of Log4j 2 appenders (File, RandomAccess File, MemoryMapped File, Console, Rewrite, other?). Use the same layout for comparison. Perhaps the PatternLayout with the {{%d \[%t\] %p %c - %m%n}} pattern.
> (-) Cost of various APIs/wrappers (SLF4J, Log4j1, JUL, Commons Logging)
> (?) Compare performance all layouts (CSV, Gelf, HTML, JSON, Pattern, RFC-5424, Serialized, Syslog, XML). Perhaps for log events with and without Throwable. TBD: any layout options to compare? (It may be good to document which features have a performance cost.)
> (?) Cost of various Pattern Layout options. Are there any converters that are particularly expensive (other than location)?
> (?) JDBC appenders? - different JDBC drivers and target databases may have very different performance. May become a big project. We could do a quick comparison of the JDBC appender to the JDK Derby DB compared against FileAppender just to get an idea of max sustained throughput?
> -------------------
> Of the existing Performance page sections:
> (-) Briefly mention that disabled logging has no measurable cost, but de-emphasize this section by moving it down the page. 
> (-) I like the part about the filters because it a) compares Log4j 2 to Logback and b) considers multithreaded applications. I'll turn this into a JMH test and show the result as a bar chart.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: log4j-dev-unsubscribe@logging.apache.org
For additional commands, e-mail: log4j-dev-help@logging.apache.org