The Boston Diaries

The ongoing saga of a programmer who doesn't live in Boston, nor does he even like Boston, but yet named his weblog/journal “The Boston Diaries.”

Go figure.

Monday, March 08, 2010

Not the Messiah

Like Handel only funnier (link via news from me). Ah, so that's what Monty Phython has been up to …

Adventures in profiling

Last month, Mark hired me on as a consultant to help him profile a program he's been writing. And while I can't describe the program (or the project the program is for), I can, however, describe the issues I've had with profiling this program.

Normally to profile a program, you compile the program using special compiler options (with gcc this is the -pg option) that instrument the program with special code to record not only how many times each function is called, but how much time is spent running each function. This information is saved when the program terminates, and you run another special program to decode this output into an easy to read format. Then you use that information to boost the performance of your program.

It's pretty straightforward.

This project, on the other hand, wasn't so straightforward. The primary issue—it's a multiprocess program; that is, it calls fork() to create child processes that do the actual work. It's the child processes that need to be profiled, but it's difficult to actually get the profile information from the child processes due to the way the profiler works (and both the GNU profiler (which I'm not allowed to use due to licensing fears) and the Sun Studio 12 profiler (which is the development platform for the project) work simularly, so this issue affects both).

Problem one—the output. The program runs and when it exits, the accumulated data is written to a file. This file is named mon.out (gmon.out if I were using GNU). In this case, the main program starts, creates several child processes. The ouput file is only generated when a process ends, and the only time a process ends in this project is when you explicitely stop the main program. This results in mon.out being overwritten by each child as they end, then overwritten again when the main process ends. So all I end up with is profile information for the main process, which tells me that the main processing loop took 99% of the time with only 0.01 seconds of CPU time (in other words—the parent process did nothing noteworthy). And there's no option in either the compiler or at runtime, to change the output file.

Or is there?

The file mon.out is generated in the current working directory of the running process. Change the working directory of the process, and the file ends up in the new working directory. So I modify the program such that each child process creates a new working directory based on the child's process ID and try again.

I ended up with one mon.out file in the working directory of the main process, and a bunch of empty directories. This leads to the second problem with profiling—you only get the output when the process calls exit() (or returns from main()), not _exit() (there is a difference between the two).

And replacing the calls to _exit() with exit() caused the program to hang (Mark even had a comment in the code about the C runtime handling fork() badly in the case of exit()).

So that pretty much killed using the compiler profiler options for this project.

Or did it?

The code is on a Sun server, which means it comes with DTrace, which is an incredible tracing facility that can be used to profile an application! Without compiling a special version of the program! Heck, you can profile any running process at any time!

It's a neat facility.

Just by using some sample scripts from the DTraceToolkit and a few examples from the Solaris Dynamic Tracing Guide, I was able to provide Mark with enough information to nearly double the performance of the program (major culprit—a ton of pointless strcpy() calls in a third party library being used, but that's about all I can say about that).

I was fortunate in that DTrace existed; between samping the program counter, recording the number of standard library calls made, and selectively checking the call stack to a few questionable calls (for instance, sprintf() calls ferror() if you can believe it, and tracking down the few hundred thousand calls to strcpy()) of selected child processes, I was able to profile this multiprocess program (and each process is multithreaded—a fact I didn't realize until later).

And if DTrace didn't exist? Well … there's the profiling equivalent to printf() debugging I could have tried …

Cache flow problems

“I keep getting these notifications that you've updated your site,” said Dad, “but when I check, I keep seeing the same page over and over again.” This was about the fourth or fifth time this topic has come up over the past few months. And each time I tell Dad to shift-reload but that doesn't seem to work for him, although the page eventually does change, after awhile.


I suspect I know the problem. Sometime in the past year I changed the configuration of my webserver to allow browsers and web proxies to better cache the content here:

ExpiresActive  On
ExpiresDefault "access plus 1 day"
ExpiresByType  image/jpeg "access plus 1 month"
ExpiresByType  image/gif  "access plus 1 month"
ExpiresByType  image/png  "access plus 1 month"
ExpiresByType  text/css   "access plus 1 month"
ExpiresByType  text/plain "access plus 1 month"

Images, style sheets, plain text files, all can be cached for a month (heck, they can probably be cached indefinitely for my blog—the content has never really changed all that much), but the web pages (in addition to the various feed files) can only be cahced for 24 hours.

And I think that's the issue.

Dad's ISP is known to aggressively cache the web (I won't name names, but it's initials are 'A', 'O' and 'L') so I may need to adjust my methods for caching values.

Obligatory Picture

[“I am NOT a number, I am … a Q-CODE!”]

Obligatory Contact Info

Obligatory Feeds

Obligatory Links

Obligatory Miscellaneous

You have my permission to link freely to any entry here. Go ahead, I won't bite. I promise.

The dates are the permanent links to that day's entries (or entry, if there is only one entry). The titles are the permanent links to that entry only. The format for the links are simple: Start with the base link for this site:, then add the date you are interested in, say 2000/08/01, so that would make the final URL:

You can also specify the entire month by leaving off the day portion. You can even select an arbitrary portion of time.

You may also note subtle shading of the links and that's intentional: the “closer” the link is (relative to the page) the “brighter” it appears. It's an experiment in using color shading to denote the distance a link is from here. If you don't notice it, don't worry; it's not all that important.

It is assumed that every brand name, slogan, corporate name, symbol, design element, et cetera mentioned in these pages is a protected and/or trademarked entity, the sole property of its owner(s), and acknowledgement of this status is implied.

Copyright © 1999-2024 by Sean Conner. All Rights Reserved.