Enhancing Java Performance with Next Generation Application Profiling

1
Enhancing Java Performance with Next Generation
Application Profiling

• Ethan Henry
• KL Group
• April 23, 1998

2
What is Profiling?

• Profiling is measuring an application,
  specifically
• where is the time being spent?
• This is classical profiling
• which method takes the most time?
• which method is called the most?
• how is memory being used?
• what kind of objects are being created?
• this in especially applicable in OO, GCed
  environments

3
Profiling vs Benchmarking

• Profiling is measuring statistics on your
  particular application
• Benchmarking is measuring the performance of a
  particular platform
• Also, optimization is an automatic technique that
  applies generic enhancements to speed up code

4
Profiling vs Benchmarking

• A profiler is not designed to tell you what VM is
  the fastest but where you application bottlenecks
  are
• Faster VMs and optimized code wont make all
  performance problems disappear

5
Why Profile?

• So, why use a profiler?
• Intuition is often wrong
• Performance is a major aspect of program
  acceptance
• Profiler advantages
• Accuracy
• Completeness
• Solid, statistical information

6
What Profiling Tells You

• Basic information
• How much time is spent in each method? (flat
  profiling)
• How many objects of each type are allocated?

7
What Profiling Tells You

• Beyond the basics
• Program flow (hierarchical profiling)
• do calls to method A cause method B to take too
  much time?
• Per-line information
• which line(s) in a given method are the most
  expensive?
• Which methods created which objects?
• Visualization aspects
• Is it easy to use the profiler to get to the
  information youre interested in?

8
How Profilers Work

• There are a variety of ways that profilers work
• There is usually a trade-off in terms of
• accuracy
• speed
• granularity of information
• intrusiveness

9
Insertion

• One technique is insertion
• Multiple flavours
• Source code insertion
• profiling code goes in with the source
• easy to do
• Object code insertion
• profiling code goes into the .o (C) or .class
  (Java) files
• hard to do

10
Insertion Pros Cons

• Insertion Pros
• can be used across a variety of platforms
• accurate (in some ways)
• Insertion Cons
• requires recompilation or relinking of the app
• profiling code may affect performance
• difficult to calculate exact impact

11
Sampling

• In sampling, the processor or VM is monitored and
  at regular intervals an interrupt executes and
  saves a snapshot of the processor state
• This data is then compared with the programs
  layout in memory to get an idea of where the
  program was at each sample

12
Sampling Pros Cons

• Sampling Pros
• no modification of app is necessary
• Sampling Cons
• a definite time/accuracy trade-off
• a high sample rate is accurate, but takes a lot
  of time
• more

13
Sampling Pros Cons

• Sampling Cons
• very small methods will almost always be missed
• if a small method is called frequently and you
  have are unlucky, small but expensive methods may
  never show up
• sampling cannot monitor memory usage

14
Instrumented VM

• Another way to collect information is to
  instrument the Java VM
• Using this technique each and every VM
  instruction can be monitored
• highly accurate

15
Instrumented VM ProsCons

• Pros
• The most accurate technique
• Can monitor memory usage data as well as time
  data
• Can easily be extended to allow remote profiling
• Cons
• The instrumented VM is platform-specific

16
JProbe Profiler

• JProbe Profiler uses an instrumented VM
• the VM is the standard Java VM, licensed from
  JavaSoft
• JProbe is accurate to the nanosecond level
• using the Pentium clock cycle timer

17
JProbe Profiler

• JProbe Profiler shows not only timer performance
  data, but also memory usage
• Memory usage is not only shown post-mortem, but
  interactively, as the app is running

18
JProbe Demo

• Time for a demo...