Profiling for the Blu-ray Java Platform

Challenges:

• Profiling frameworks like JVMTI are not available for the Blu-ray Java platform.
• Using System.out may not be feasible for some players are known to hinder the application performance when it's used.
• The system clock on the players is too slow to measure anything meaningful on non-PS3 players.

Solution:

Our team at Sun have developed an innovative way of profiling of Blu-ray Java applications. I will refer to it as hdcookbook profiler as it's made available through our open source project at: https://hdcookbook.dev.java.net

More about hdcookbook profiler:

• The execution time is measured using PC's clock, not the player's clock. This enables logging of nanoseconds instead of milliseconds.
  
• The profiling data is collected with minimum heap data and time overhead.
  
• The profiler does not generate any profiling information on its own. This is totally under user's control. The user explcitly adds the profiling points of interest to the application.
  
• The profiling events (start/stop timer) are communicated to the PC by means of UDP packets. The profiling events can be debug messages as well.
  
• For viewing the profiled data, we have two UI applications available, one of them was developed for immediate use and the second one came out later. A screen shot of the ProfileBrowser (second one) is shown here:

  
  

  
  

  Browser:

  • The vertical bars represent the execution times for each method. The RED stretches indicate the execution times that are out side the standard window of execution times for a given method. If you see many red colored stretches, you can increase the deviation factor say by: 2 ( i.e 2 * standard deviation) or more to identify execution times that are taking way too longer than the standard execution times for that method.
    
  • You can select the time unit that is convenient for analyzing the data at hand. We have set Microseconds as a default time unit. You can switch between micro, nano, millis and secs without resulting in any loss of precision during switching.
    
  • The range slider (right side bar; looks like scroll bar but with arrows pointing in opposite directions) allows zoom-in and zoom-out (narrowing down) of the time scale with up and down mouse dragging. That means you could zoom into a time range of interest by just dragging the mouse.
    
  • Method Filtering: You can provide an expression to filter the data on the display (the editable box is on the left at the bottom) and get the timeline for selected method/s. For example, you could type in an expression: a | b This will only plot methods names starting with letters a or b. This is useful for focussing on the individual method/s of interest to see how the timeline looks for them.
  
  Also see this write up about profiler

  Posted at 04:37AM Aug 03, 2009 by jaya_hangal in Java  |  Comments[1]