Performance Tools for Optimizing Linux: Process-Specific CPU

This chapter is from the book 

4.2 The Tools

Linux has a variety of tools to help you determine which pieces of an application are the primary users of the CPU. This section describes these tools.

4.2.1 time

The time command performs a basic function when testing a command’s performance, yet it is often the first place to turn. The time command acts as a stopwatch and times how long a command takes to execute. It measures three types of time. First, it measures the real or elapsed time, which is the amount of time between when the program started and finished execution. Next, it measures the user time, which is the amount of time that the CPU spent executing application code on behalf of the program. Finally, time measures system time, which is the amount of time the CPU spent executing system or kernel code on behalf of the application.

4.2.1.1 CPU Performance-Related Options

The time command (see Table 4-1) is invoked in the following manner:

time [-v] application

The application is timed, and information about its CPU usage is displayed on standard output after it has completed.

Table 4-1 time Command-Line Options

Option	Explanation
-v	This option presents a verbose display of the program’s time and statistics. Some statistics are zeroed out, but more statistics are valid with Linux kernel v2.6 than with Linux kernel v2.4.
	Most of the valid statistics are present in both the standard and verbose mode, but the verbose mode provides a better description for each statistic.

Table 4-2 describes the valid output statistic that the time command provides. The rest are not measured and always display zero.

Table 4-2 CPU-Specific time Output

Column	Explanation
User time (seconds)	This is the number of seconds of CPU spent by the application.
System time (seconds)	This is the number of seconds spent in the Linux kernel on behalf of the application.
Elapsed (wall-clock) time (h:mm:ss or m:ss)	This is the amount of time elapsed (in wall-clock time) between when the application was launched and when it completed.
Percent of CPU this job got	This is the percentage of the CPU that the process consumed as it was running.
Major (requiring I/O) page faults	The number of major page faults or those that required a page of memory to be read from disk.
Minor (reclaiming a frame) page faults	The number of minor page faults or those that could be filled without going to disk.
Swaps	This is the number of times the process was swapped to disk.
Voluntary context switches	The number of times the process yielded the CPU (for example, by going to sleep).
Involuntary context switches:	The number of times the CPU was taken from the process.
Page size (bytes)	The page size of the system.
Exit status	The exit status of the application.

This command is a good way to start an investigation. It displays how long the application is taking to execute and how much of that time is spent in the Linux kernel versus your application.

4.2.1.2 Example Usage

The time command included on Linux is a part of the cross-platform GNU tools. The default command output prints a host of statistics about the commands run, even if Linux does not support them. If the statistics are not available, time just prints a zero. The following command is a simple invocation of the time command. You can see in Listing 4.1 that the elapsed time (~3 seconds) is much greater than the sum of the user (0.9 seconds) and system (0.13 seconds) time, because the application spends most of its time waiting for input and little time using the processor.

Listing 4.1

[ezolt@wintermute manuscript]$ /usr/bin/time gcalctool
0.91user 0.13system 0:03.37elapsed 30%CPU (0avgtext+0avgdata 0maxresident)k
0inputs+0outputs (2085major+369minor)pagefaults 0swaps

Listing 4.2 is an example of time displaying verbose output. As you can see, this output shows much more than the typical output of time. Unfortunately, most of the statistics are zeros, because they are not supported on Linux. For the most part, the information provided in verbose mode is identical to the output provided in standard mode, but the statistics’ labels are much more descriptive. In this case, we can see that this process used 15 percent of the CPU when it was running, and spent 1.15 seconds running user code with .12 seconds running kernel code. It accrued 2,087 major page faults, or memory faults that did not require a trip to disk; it accrued 371 page faults that did require a trip to disk. A high number of major faults would indicate that the operating system was constantly going to disk when the application tried to use memory, which most likely means that the kernel was swapping a significant amount.

Listing 4.2

[ezolt@wintermute manuscript]$ /usr/bin/time --verbose gcalctool
    Command being timed: "gcalctool"
    User time (seconds): 1.15
    System time (seconds): 0.12
    Percent of CPU this job got: 15%
    Elapsed (wall clock) time (h:mm:ss or m:ss): 0:08.02
    Average shared text size (kbytes): 0
    Average unshared data size (kbytes): 0
    Average stack size (kbytes): 0
    Average total size (kbytes): 0
    Maximum resident set size (kbytes): 0
    Average resident set size (kbytes): 0
    Major (requiring I/O) page faults: 2087
    Minor (reclaiming a frame) page faults: 371
    Voluntary context switches: 0
    Involuntary context switches: 0
    Swaps: 0
    File system inputs: 0
    File system outputs: 0
    Socket messages sent: 0
    Socket messages received: 0
    Signals delivered: 0
    Page size (bytes): 4096
    Exit status: 0

Note that the bash shell has a built-in time command, so if you are running bash and execute time without specifying the path to the executable, you get the following output:

[ezolt@wintermute manuscript]$ time gcalctool
 
real  0m3.409s
user  0m0.960s
sys   0m0.090s

The bash built-in time command can be useful, but it provides a subset of the process execution information.

4.2.2 strace

strace is a tool that traces the system calls that a program makes while executing. System calls are function calls made into the Linux kernel by or on behalf of an application. strace can show the exact system calls that were made and proves incredibly useful to determine how an application is using the Linux kernel. Tracing down the frequency and length of system calls can be especially valuable when analyzing a large program or one you do not understand completely. By looking at the strace output, you can get a feel for how the application is using the kernel and what type of functions it depends on.

strace can also be useful when you completely understand an application, but if that application makes calls to system libraries (such as libc or GTK.) In this case, even though you know where the application makes every system call, the libraries might be making more system calls on behalf of your application. strace can quickly show you what calls these libraries are making.

Although strace is mainly intended to trace the interaction of processes with the kernel by showing the arguments and results for every system call an application makes, strace can also provide summary information that is a little less daunting. After the run of an application, strace can provide a table showing the frequency of each system call and the total time spent in calls of that type. This table can be a crucial first piece of information in understanding how your program is interacting with the Linux kernel.

4.2.2.1 CPU Performance-Related Options

The following invocation of strace is most useful for performance testing:

strace [-c] [-p pid] [-o file] [--help] [ command [ arg ... ]]

If strace is run without any options, it displays all the system calls made by the given command on standard error. This can be helpful when trying to figure out why an application is spending a large amount of time in the kernel. Table 4-3 describes a few strace options that are also helpful when tracing a performance problem.

Table 4-3 strace Command-Line Options

Option	Explanation
-c	This causes strace to print out a summary of statistics rather than an individual list of all the system calls that are made.
-p pid	This attaches to the process with the given PID and starts tracing.
-o file	The output of strace will be saved in file.
--help	Lists a complete summary of the strace options

Table 4-4 explains the statistics present in output of the strace summary option. Each line of output describes a set of statistics for a particular system call.

Table 4-4 CPU-Specific strace Output

Column	Explanation
% time	Of the total time spent making system calls, this is the percentage of time spent on this one.
seconds	This the total number of seconds spent in this system call.
usecs/call	This is the number of microseconds spent per system call of this type.
calls	This is the total number of calls of this type.
errors	This is the number of times that this system call returned an error.

Although the options just described are most relevant to a performance investigation, strace can also filter the types of system calls that it traces. The options to select the system calls to trace are described in detail with the --help option and in the strace man page. For general performance tuning, it is usually not necessary to use them; if needed, however, they exist.

4.2.2.2 Example Usage

Listing 4.3 is an example of using strace to gather statistics about which system calls an application is making. As you can see, strace provides a nice profile of the system’s calls made on behalf of an application, which, in this case, is oowriter. In this example, we look at how oowriter is using the read system call. We can see that read is taking the 20 percent of the time by consuming a total of 0.44 seconds. It is called 2,427 times and, on average, each call takes 184 microseconds. Of those calls, 26 return an error.

Listing 4.3

[ezolt@wintermute tmp]$ strace -c oowriter
execve("/usr/bin/oowriter", ["oowriter"], [/* 35 vars */]) = 0
Starting OpenOffice.org ...

% time   seconds    usecs/call   calls    errors      syscall
------ ----------- ----------- --------- --------- ----------------
 20.57  0.445636           184      2427        26 read
 18.25  0.395386           229      1727           write
 11.69  0.253217           338       750       514 access
 10.81  0.234119         16723        14         6 waitpid
  9.53  0.206461          1043       198           select
  4.73  0.102520           201       511        55 stat64
  4.58  0.099290          154        646           gettimeofday
  4.41  0.095495           58       1656        15 lstat64
  2.51  0.054279          277        196           munmap
  2.32  0.050333          123        408           close
  2.07  0.044863           66        681       297 open
  1.98  0.042879          997        43            writev
  1.18  0.025614           12      2107            lseek
  0.95  0.020563         1210        17            unlink
  0.67  0.014550          231        63            getdents64
  0.58  0.012656           44       286            mmap2
  0.53  0.011399           68       167          2 ioctl
  0.50  0.010776          203        53            readv
  0.44  0.009500         2375         4          3 mkdir
  0.33  0.007233          603        12            clone
  0.29  0.006255           28       224            old_mmap
  0.24  0.005240         2620         2            vfork
  0.24  0.005173           50       104            rt_sigprocmask
  0.11  0.002311           8        295            fstat64

strace does a good job of tracking a process, but it does introduce some overhead when it is running on an application. As a result, the number of calls that strace reports is probably more reliable than the amount of time that it reports for each call. Use the times provided by strace as a starting point for investigation rather than a highly accurate measurement of how much time was spent in each call.

4.2.3 ltrace

ltrace is similar in concept to strace, but it traces the calls that an application makes to libraries rather than to the kernel. Although it is primarily used to provide an exact trace of the arguments and return values of library calls, you can also use ltrace to summarize how much time was spent in each call. This enables you to figure out both what library calls the application is making and how long each is taking.

Be careful when using ltrace because it can generate misleading results. The time spent may be counted twice if one library function calls another. For example, if library function foo() calls function bar(), the time reported for function foo() will be all the time spent running the code in function foo() plus all the time spent in function bar().

With this caveat in mind, it still is a useful tool to figure out how an application is behaving.

4.2.3.1 CPU Performance-Related Options

ltrace provides similar functionality to strace and is invoked in a similar way:

ltrace [-c] [-p pid] [-o filename] [-S] [--help] command

In the preceding invocation, command is the command that you want ltrace to trace. Without any options to ltrace, it displays all the library calls to standard error. Table 4-5 describes the ltrace options that are most relevant to a performance investigation.

Table 4-5 ltrace Command-Line Options

Option	Explanation
-c	This option causes ltrace to print a summary of all the calls after the command has completed.
-S	ltrace traces system calls in addition to library calls, which is identical to the functionality strace provides.
-p pid	This traces the process with the given PID.
-o file	The output of strace is saved in file.
--help	Displays help information about ltrace.

Again, the summary mode provides performance statistics about the library calls made during an application’s execution. Table 4-6 describes the meanings of these statistics.

Table 4-6 CPU-Specific ltrace Output

Column	Explanation
% time	Of the total time spent making library calls, this is the percentage of time spent on this one.
seconds	This is the total number of seconds spent in this library call.
usecs/call	This is the number of microseconds spent per library call of this type.
calls	This is the total number of calls of this type.
function	This is the name of the library call.

Much like strace, ltrace has a large number of options that can modify the functions that it traces. These options are described by the ltrace --help command and in detail in the ltrace man page.

4.2.3.2 Example Usage

Listing 4.4 is a simple example of ltrace running on the xeyes command. xeyes is an X Window application that pops up a pair of eyes that follow your mouse pointer around the screen.

Listing 4.4

[ezolt@localhost manuscript]$ ltrace -c /usr/X11R6/bin/xeyes
% time   seconds    usecs/call   calls       function
------ ----------- ----------- --------- --------------------
 18.65  0.065967         65967         1 XSetWMProtocols
 17.19  0.060803            86       702 hypot
 12.06  0.042654           367       116 XQueryPointer
  9.51  0.033632         33632         1 XtAppInitialize
  8.39  0.029684            84       353 XFillArc
  7.13  0.025204           107       234 cos
  6.24  0.022091            94       234 atan2
  5.56  0.019656            84       234 sin
  4.62  0.016337           139       117 XtAppAddTimeOut
  3.19  0.011297            95       118 XtWidgetToApplicationContext
  3.06  0.010827            91       118 XtWindowOfObject
  1.39  0.004934          4934         1 XtRealizeWidget
  1.39  0.004908          2454         2 XCreateBitmapFromData
  0.65  0.002291          2291         1 XtCreateManagedWidget
  0.12  0.000429           429         1 XShapeQueryExtension
  0.09  0.000332           332         1 XInternAtom
  0.09  0.000327            81         4 XtDisplay
  0.09  0.000320           106         3 XtGetGC
  0.05  0.000168            84         2 XSetForeground
  0.05  0.000166            83         2 XtScreen
  0.04  0.000153           153         1 XtParseTranslationTable
  0.04  0.000138           138         1 XtSetValues
  0.04  0.000129           129         1 XmuCvtStringToBackingStore
  0.03  0.000120           120         1 XtDestroyApplicationContext
  0.03  0.000116           116         1 XtAppAddActions
  0.03  0.000109           109         1 XCreatePixmap
  0.03  0.000108           108         1 XtSetLanguageProc
  0.03  0.000104           104         1 XtOverrideTranslations
  0.03  0.000102           102         1 XtWindow
  0.03  0.000096            96         1 XtAddConverter
  0.03  0.000093            93         1 XtCreateWindow
  0.03  0.000093            93         1 XFillRectangle
  0.03  0.000089            89         1 XCreateGC
  0.03  0.000089            89         1 XShapeCombineMask
  0.02  0.000087            87         1 XclearWindow
  0.02  0.000086            86         1 XFreePixmap
------ ----------- ----------- --------- --------------------
100.00  0.353739                    2261 total

In Listing 4.4, the library functions XSetWMProtocols, hypot, and XQueryPointer take 18.65 percent, 17.19 percent, and 12.06 percent of the total time spent in libraries. The call to the second most time-consuming function, hypot, is made 702 times, and the call to most time-consuming function, XSetWMProtocols, is made only once. Unless our application can completely remove the call to XSetWMProtocols, we are likely stuck with whatever time it takes. It is best to turn our attention to hypot. Each call to this function is relatively lightweight; so if we can reduce the number of times that it is called, we may be able to speed up the application. hypot would probably be the first function to be investigated if the xeyes application was a performance problem. Initially, we would determine what hypot does, but it is unclear where it may be documented. Possibly, we could figure out which library hypot belongs to and read the documentation for that library. In this case, we do not have to find the library first, because a man page exists for the hypot function. Running man hypot tells us that the hypot function will calculate the distance (hypotenuse) between two points and is part of the math library, libm. However, functions in libraries may have no man pages, so we would need to be able to determine what library a function is part of without them. Unfortunately, ltrace does not make it at obvious which library a function is from. To figure it out, we have to use the Linux tools ldd and objdump. First, ldd is used to display which libraries are used by a dynamically linked application. Then, objdump is used to search each of those libraries for the given function. In Listing 4.5, we use ldd to see which libraries are used by the xeyes application.

Listing 4.5

[ezolt@localhost manuscript]$ ldd /usr/X11R6/bin/xeyes
    linux-gate.so.1 => (0x00ed3000)
    libXmu.so.6 => /usr/X11R6/lib/libXmu.so.6 (0x00cd4000)
    libXt.so.6 => /usr/X11R6/lib/libXt.so.6 (0x00a17000)
    libSM.so.6 => /usr/X11R6/lib/libSM.so.6 (0x00368000)
    libICE.so.6 => /usr/X11R6/lib/libICE.so.6 (0x0034f000)
    libXext.so.6 => /usr/X11R6/lib/libXext.so.6 (0x0032c000)
    libX11.so.6 => /usr/X11R6/lib/libX11.so.6 (0x00262000)
    libm.so.6 => /lib/tls/libm.so.6 (0x00237000)
    libc.so.6 => /lib/tls/libc.so.6 (0x0011a000)
    libdl.so.2 => /lib/libdl.so.2 (0x0025c000)
    /lib/ld-linux.so.2 => /lib/ld-linux.so.2 (0x00101000)

Now that the ldd command has shown the libraries that xeyes uses, we can use the objdump command to figure out which library the function is in. In Listing 4.6, we look for the hypot symbol in each of the libraries that xeyes is linked to. The -T option of objdump lists all the symbols (mostly functions) that the library relies on or provides. By using fgrep to look at output lines that have .text in it, we can see which libraries export the hypot function. In this case, we can see that the libm library is the only library that contains the hypot function.

Listing 4.6

[/tmp]$ objdump -T /usr/X11R6/lib/libXmu.so.6  | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /usr/X11R6/lib/libXt.so.6   | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /usr/X11R6/lib/libSM.so.6   | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /usr/X11R6/lib/libICE.so.6  | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /usr/X11R6/lib/libXext.so.6 | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /usr/X11R6/lib/libX11.so.6  | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /lib/tls/libm.so.6          | fgrep ".text" | grep "hypot"
00247520 w  DF .text 000000a9 GLIBC_2.0        hypotf
0024e810 w  DF .text 00000097 GLIBC_2.0        hypotl
002407c0 w  DF .text 00000097 GLIBC_2.0        hypot
[/tmp]$ objdump -T /lib/tls/libc.so.6          | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /lib/libdl.so.2             | fgrep ".text" | grep "hypot"
[/tmp]$ objdump -T /lib/ld-linux.so.2          | fgrep ".text" | grep "hypot"

The next step might be to look through the source of xeyes to figure out where hypot is called and, if possible, reduce the number of calls made to it. An alternative solution is to look at the source of hypot and try to optimize the source code of the library.

By enabling you to investigate which library calls are taking a long time to complete, ltrace enables you to determine the cost of each library call that an application makes.

4.2.4 ps (Process Status)

ps is an excellent command to track a process’s behavior as it runs.

It provides detailed static and dynamic statistics about currently running processes. ps provides static information, such as command name and PID, as well as dynamic information, such as current use of memory and CPU.

4.2.4.1 CPU Performance-Related Options

ps has many different options and can retrieve many different statistics about the state of a running application. The following invocations are those options most related to CPU performance and will show information about the given PID:

ps [-o etime,time,pcpu,command] [-u user] [-U user] [PID]

The command ps is probably one of the oldest and feature-rich commands to extract performance information, which can make its use overwhelming. By only looking at a subset of the total functionality, it is much more manageable. Table 4-7 contains the options that are most relevant to CPU performance.

Table 4-7 ps Command-Line Options

Option	Explanation
-o <statistic>	This option enables you to specify exactly what process statistics you want to track. The different statistics are specified in a comma-separated list with no spaces.
etime	Statistic: Elapsed time is the amount of time since the program began execution.
time	Statistic: CPU time is the amount of system plus user time the process spent running on the CPU.
pcpu	Statistic: The percentage of CPU that the process is currently consuming.
command	Statistic: This is the command name.
-A	Shows statistics about all processes.
-u user	Shows statistics about all processes with this effective user ID.
-U user	Shows statistic about all processes with this user ID.

ps provides myriad different performance statistics in addition to CPU statistics, many of which, such as a process’s memory usage, are discussed in subsequent chapters.

4.2.4.2 Example Usage

This example shows a test application that is consuming 88 percent of the CPU and has been running for 6 seconds, but has only consumed 5 seconds of CPU time:

[ezolt@wintermute tmp]$ ps -o etime,time,pcpu,cmd 10882
  ELAPSED   TIME %CPU CMD
   00:06 00:00:05 88.0 ./burn

In Listing 4.7, instead of investigating the CPU performance of a particular process, we look at all the processes that a particular user is running. This may reveal information about the amount of resources a particular user consumes. In this case, we look at all the processes that the netdump user is running. Fortunately, netdump is a tame user and is only running bash, which is not taking up any of the CPU, and top, which is only taking up 0.5 percent of the CPU.

Listing 4.7

[/tmp]$ ps -o time,pcpu,command -u netdump
  TIME %CPU COMMAND
00:00:00 0.0 -bash
00:00:00 0.5 top

Unlike time, ps enables us to monitor information about a process currently running. For long-running jobs, you can use ps to periodically check the status of the process (instead of using it only to provide statistics about the program’s execution after it has completed).

4.2.5 ld.so (Dynamic Loader)

When a dynamically linked application is executed, the Linux loader, ld.so, runs first. ld.so loads all the application’s libraries and connects symbols that the application uses with the functions the libraries provide. Because different libraries were originally linked at different and possibly overlapping places in memory, the linker needs to sort through all the symbols and make sure that each lives at a different place in memory. When a symbol is moved from one virtual address to another, this is called a relocation. It takes time for the loader to do this, and it is much better if it does not need to be done at all. The prelink application aims to do that by rearranging the system libraries of the entire systems so that they do not overlap. An application with a high number of relocations may not have been prelinked.

The Linux loader usually runs without any intervention from the user, and by just executing a dynamic program, it is run automatically. Although the execution of the loader is hidden from the user, it still takes time to run and can potentially slow down an application’s startup time. When you ask for loader statistics, the loader shows the amount of work it is doing and enables you to figure out whether it is a bottleneck.

4.2.5.1 CPU Performance-Related Options

The ld command is invisibly run for every Linux application that uses shared libraries. By setting the appropriate environment variables, we can ask it to dump information about its execution. The following invocation influences ld execution:

env LD_DEBUG=statistics,help LD_DEBUG_OUTPUT=filename <command>

The debugging capabilities of the loader are completely controlled with environmental variables. Table 4-8 describes these variables.

Table 4-8 ld Environmental Variables

Option	Explanation
LD_DEBUG=statistics	This turns on the display of statistics for ld.
LD_DEBUG=help	This displays information about the available debugging statistics.

Table 4-9 describes some of the statistics that ld.so can provide. Time is given in clock cycles. To convert this to wall time, you must divide by the processor’s clock speed. (This information is available from cat /proc/cpuinfo.)

Table 4-9 CPU Specific ld.so Output

Column	Explanation
total startup time in dynamic loader	The total amount of time (in clock cycles) spent in the load before the application started to execute.
time needed for relocation	The total amount of time (in clock cycles) spent relocating symbols.
number of relocations	The number of new relocation calculations done before the application’s execution began.
number of relocations from cache	The number of relocations that were precalculated and used before the application started to execute.
number of relative relocations	The number of relative relocations.
time needed to load objects	The time needed to load all the libraries that an application is using.
final number of relocations	The total number of relocations made during an application run (including those made by dlopen).
final number of relocations from cache	The total number of relocations that were precalculated.

The information provided by ld can prove helpful in determining how much time is being spent setting up dynamic libraries before an application begins executing.

4.2.5.2 Example Usage

In Listing 4.8, we run an application with the ld debugging environmental variables defined. The output statistics are saved into the lddebug file. Notice that the loader shows two different sets of statistics. The first shows all the relocations that happen during startup, whereas the last shows all the statistics after the program closes. These can be different values if the application uses functions such as dlopen, which allows shared libraries to be mapped into an application after it has started to execute. In this case, we see that 83 percent of the time spent in the loader was doing allocations. If the application had been prelinked, this would have dropped close to zero.

Listing 4.8

 [ezolt@wintermute ezolt]$ env LD_DEBUG=statistics LD_DEBUG_OUTPUT=lddebug gcalctool
 [ezolt@wintermute ezolt]$ cat lddebug.2647
   2647:
   2647:   runtime linker statistics:
   2647:    total startup time in dynamic loader: 40820767 clock cycles
   2647:         time needed for relocation: 33896920 clock cycles (83.0%)
   2647:           number of relocations: 2821
   2647:      number of relocations from cache: 2284
   2647:       number of relative relocations: 27717
   2647:        time needed to load objects: 6421031 clock cycles (15.7%)
   2647:
   2647:   runtime linker statistics:
   2647:        final number of relocations: 6693
   2647:   final number of relocations from cache: 2284

If ld is determined to be the cause of sluggish startup time, it may be possible to reduce the amount of startup time by pruning the number of libraries that an application relies on or running prelink on the system.

4.2.6 gprof

A powerful way to profile applications on Linux is to use the gprof profiling command. gprof can show the call graph of application and sample where the application time is spent. gprof works by first instrumenting your application and then running the application to generate a sample file. gprof is very powerful, but requires application source and adds instrumentation overhead. Although it can accurately determine the number of times a function is called and approximate the amount of time spent in a function, the gprof instrumentation will likely change the timing characteristics of the application and slow down its execution.

4.2.6.1 CPU Performance-Related Options

To profile an application with gprof, you must have access to application source. You must then compile that application with a gcc command similar to the following:

gcc -gp -g3 -o app app.c 

First, you must compile the application with profiling turned on, using gcc’s -gp option. You must take care not to strip the executable, and it is even more helpful to turn on symbols when compiling using the -g3 option. Symbol information is necessary to use the source annotation feature of gprof. When you run your instrumented application, an output file is generated. You can then use the gprof command to display the results. The gprof command is invoked as follows:

gprof [-p –flat-profile -q --graph --brief -A –annotated-source ] app

The options described in Table 4-10 specify what information gprof displays.

Table 4-10 gprof Command-Line Options

Option	Explanation
--brief	This option abbreviates the output of gprof. By default, gprof prints out all the performance information and a legend to describe what each metric means. This suppresses the legend.
-p or --flat-profile	This option prints out the total amount of time spent and the number of calls to each function in the application.
-q or --graph	This option prints out a call graph of the profiled application. This shows how the functions in the program called each other, how much time was spent in each of the functions, and how much was spent in the functions of the children.
-A or --annotated-source	This shows the profiling information next to the original source code.

Not all the output statistics are available for a particular profile. Which output statistic is available depends on how the application was compiled for profiling.

4.2.6.2 Example Usage

When profiling an application with gprof, the first step is to compile the application with profiling information. The compiler (gcc) inserts profiling information into the application and, when the application is run, it is saved into a file named gmon.out. The burn test application is fairly simple. It clears a large area of memory and then calls two functions, a() and b(), which each touch this memory. Function a() touches the memory 10 times as often as function b().

First, we compile the application:

[ezolt@wintermute test_app]$ gcc -pg -g3 -o burn_gprof burn.c

After we run in it, we can analyze the output. This is shown in Listing 4.9.

Listing 4.9

[ezolt@wintermute test_app]$ gprof --brief -p ./burn_gprof
Flat profile:
 
Each sample counts as 0.01 seconds.
 %  cumulative  self       self   total
 time  seconds  seconds  calls  s/call  s/call name
 91.01    5.06     5.06      1    5.06    5.06 a
  8.99    5.56     0.50      1    0.50    0.50 b

In Listing 4.9, you can see gprof telling us what we already knew about the application. It has two functions, a() and b(). Each function is called once, and a() takes 10 times (91 percent) the amount of time to complete than b() (8.99 percent). 5.06 seconds of time is spent in the function a(), and .5 seconds is spent in function b().

Listing 4.10 shows the call graph for the test application. The <spontaneous> comment listed in the output means that although gprof did not record any samples in main(), it deduced that main() must have been run, because functions a() and b() both had samples, and main was the only function in the code that called them. gprof most likely did not record any samples in main() because it is a very short function.

Listing 4.10

   [ezolt@wintermute test_app]$ gprof --brief -q ./burn_gprof
            Call graph
 
 
granularity: each sample hit covers 4 byte(s) for 0.18% of 5.56 seconds
 
index % time  self  children  called   name
                                           <spontaneous>
[1]    100.0  0.00    5.56           main [1]
              5.06    0.00     1/1      a [2]
              0.50    0.00     1/1      b [3]
-----------------------------------------------
              5.06    0.00     1/1          main [1]
[2]     91.0  5.06    0.00     1       a [2]
-----------------------------------------------
              0.50    0.00    1/1      main [1]
[3]      9.0  0.50    0.00    1        b [3]
-----------------------------------------------

Index by function name
 
  [2] a            [3] b

Finally, gprof can annotate the source code to show how often each function is called. Notice that Listing 4.11 does not show the time spent in the function; instead, it shows the number of times the function was called. As shown in the previous examples for gprof, a() actually took 10 times as long as b(), so be careful when optimizing. Do not assume that functions that are called a large number of times are actually using the CPU for a large amount of time or that functions that are called a few number of times are necessarily taking a small amount of the CPU time.

Listing 4.11

[ezolt@wintermute test_app]$ gprof -A burn_gprof
*** File /usr/src/perf/process_specific/test_app/burn.c:
        #include <string.h>
         
        #define ITER 10000
        #define SIZE 10000000
        #define STRIDE 10000
         
        char test[SIZE];
         
        void a(void)
      1 -> {
         int i=0,j=0;
         for (j=0;j<10*ITER ; j++)
          for (i=0;i<SIZE;i=i+STRIDE)
           {
            test[i]++;
           }
        }
         
        void b(void)
      1 -> {
         int i=0,j=0;
         for (j=0;j<ITER; j++)
          for (i=0;i<SIZE;i=i+STRIDE)
           {
            test[i]++;
           }
        }
         
         
        main()
    ##### -> {
          
         /* Arbitrary value*/
         memset(test, 42, SIZE);
         a();
         b();
        }
 
 
Top 10 Lines:
 
   Line   Count
 
    10     1
    20     1
 
Execution Summary:
 
    3  Executable lines in this file
    3  Lines executed
  100.00  Percent of the file executed
 
    2  Total number of line executions
   0.67  Average executions per line

gprof provides a good summary of how many times functions or source lines in an application have been run and how long they took.

4.2.7 oprofile (II)

As discussed in Chapter 2, "Performance Tools: System CPU," you can use oprofile to track down the location of different events in the system or an application. oprofile is a lower-overhead tool than gprof. Unlike gprof, it does not require an application to be recompiled to be used. oprofile can also measure events not supported by gprof. Currently, oprofile can only support call graphs like those that gprof can generate with a kernel patch, whereas gprof can run on any Linux kernel.

4.2.7.1 CPU Performance-Related Options

The oprofile discussion in the section, "System-Wide Performance Tools" in Chapter 2 covers how to start profiling with oprofile. This section describes the parts of oprofile used to analyze the results of process-level sampling.

oprofile has a series of tools that display samples that have been collected. The first tool, opreport, displays information about how samples are distributed to the functions within executables and libraries. It is invoked as follows:

opreport [-d --details -f --long-filenames -l --symbols -l] application

Table 4-11 describes a few commands that can modify the level of information that opreport provides.

Table 4-11 opreport Command-Line Options

Option	Explanation
-d or --details	This shows an instruction-level breakdown of all the collected samples.
-f or --long-filenames	This shows the complete path name of the application being analyzed.
-l or --symbols	This shows how an application’s samples are distributed to its symbols. This enables you to see what functions have the most samples attributed to them.

The next command that you can use to extract information about performance samples is opannotate. opannotate can attribute samples to specific source lines or assembly instructions. It is invoked as follows:

opannotate [-a --assembly] [-s --source] application

The options described in Table 4-12 enable you to specify exactly what information opannotate will provide. One word of caution: because of limitations in the processor hardware counters at the source line and instruction level, the sample attributions may not be on the exact line that caused them; however, they will be near the actual event.

Table 4-12 opannotate Command-Line Options

Option	Explanation
-s or --source --	This shows the collected samples next to the application’s source code.
-a or --assembly	This shows the samples collected next to the assembly code of the application.
-s and -a	If both -s and -a are specified, opannotate intermingles the source and assembly code with the samples.

When using opannotate and opreport, it is always best to specify the full path name to the application. If you do not, you may receive a cryptic error message (if oprofile cannot find the application’s samples). By default, when displaying results, opreport only shows the executable name, which can be ambiguous in a system with multiple executables or libraries with identical names. Always specify the -f option so that opreport shows the complete path to the application.

oprofile also provides a command, opgprof, that can export the samples collected by oprofile into a form that gprof can digest. It is invoked in the following way:

opgprof application

This command takes the samples of application and generates a gprof-compatible profile. You can then view this file with the gprof command.

4.2.7.2 Example Usage

Because we already looked at oprofile in the section, "System-Wide Performance Tools" in Chapter 2, the examples here show you how to use oprofile to track down a performance problem to a particular line of source code. This section assumes that you have already started the profiling using the opcontrol command. The next step is to run the program that is having the performance problem. In this case, we use the burn program, which is the same that we used in the gprof example. We start our test program as follows:

 [ezolt@wintermute tmp]$ ./burn

After the program finishes, we must dump oprofile’s buffers to disk, or else the samples will not be available to opreport. We do that using the following command:

[ezolt@wintermute tmp]$ sudo opcontrol -d

Next, in Listing 4.12 we ask opreport to tell us about the samples relevant to our test application, /tmp/burn. This gives us an overall view of how many cycles our application consumed. In this case, we see that 9,939 samples were taken for our application. As we dig into the oprofile tools, we will see how these samples are distributed within the burn application.

Listing 4.12

[ezolt@wintermute tmp]$ opreport -f /tmp/burn
CPU: PIII, speed 467.731 MHz (estimated)
Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with a 
unit mask of 0x00 (No unit mask) count 233865
   9939 100.0000 /tmp/burn

Next, in Listing 4.13, we want to see which functions in the burn application had all the samples. Because we are using the CPU_CLK_UNHALTED event, this roughly corresponds to the relative amount of time spent in each function. By looking at the output, we can see that 91 percent of the time was spent in function a(), and 9 percent was spent in function b().

Listing 4.13

[ezolt@wintermute tmp]$ opreport -l /tmp/burn
CPU: PIII, speed 467.731 MHz (estimated)
Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with a 
unit mask of 0x00 (No unit mask) count 233865
vma   samples %      symbol name
08048348 9033   90.9118   a
0804839e 903    9.0882   b

In Listing 4.14, we ask opreport to show us which virtual addresses have samples attributed to them. In this case, it appears as if the instruction at address 0x0804838a has 75 percent of the samples. However, it is currently unclear what this instruction is doing or why.

Listing 4.14

[ezolt@wintermute tmp]$ opreport -d /tmp/burn
CPU: PIII, speed 467.731 MHz (estimated)
Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with a 
unit mask of 0x00 (No unit mask) count 233865
vma   samples %      symbol name
08048348 9033   90.9118   a
 08048363 4     0.0443
 08048375 431    4.7714
 0804837c 271    3.0001
 0804837e 1     0.0111
 08048380 422    4.6718
 0804838a 6786   75.1245
 08048393 1114   12.3326
 08048395 4     0.0443
0804839e 903    9.0882   b
 080483cb 38    4.2082
 080483d2 19    2.1041
 080483d6 50    5.5371
 080483e0 697   77.1872
 080483e9 99    10.9635

Generally, it is more useful to know the source line that is using all the CPU time rather than the virtual address of the instruction that is using it. It is not always easy to figure out the correspondence between a source line and a particular instruction; so, in Listing 4.15, we ask opannotate to do the hard work and show us the samples relative to the original source code (rather than the instruction's virtual address).

Listing 4.15

[ezolt@wintermute tmp]$ opannotate --source /tmp/burn
/*
 * Command line: opannotate --source /tmp/burn
 *
 * Interpretation of command line:
 * Output annotated source file with samples
 * Output all files
 *
 * CPU: PIII, speed 467.731 MHz (estimated)
 * Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with 
 a unit mask of 0x00 (No unit mask) count 233865
 */
/*
 * Total samples for file : "/tmp/burn.c"
 *
 *  9936 100.0000
 */
 
 
        :#include <string.h>
        :
        :#define ITER 10000
        :#define SIZE 10000000
        :#define STRIDE 10000
        :
        :char test[SIZE];
        :
        :void a(void)
        :{ /* a total:  9033 90.9118 */
        : int i=0,j=0;
   8 0.0805 : for (j=0;j<10*ITER ; j++)
 8603 86.5841 :  for (i=0;i<SIZE;i=i+STRIDE)
        :   {
  422 4.2472 :    test[i]++;
        :   }
        :}
        :
        :void b(void)
        :{ /* b total:  903 9.0882 */
        : int i=0,j=0;
        : for (j=0;j<ITER; j++)
  853 8.5849 :  for (i=0;i<SIZE;i=i+STRIDE)
        :   {
  50 0.5032 :    test[i]++;
        :   }
        :}
        :
        :
        :main()
        :{
        :
        : /* Arbitrary value*/
        : memset(test, 42, SIZE);
        : a();
        : b();
        :}

As you can see in Listing 4.15, opannotate attributes most of the samples (86.59 percent) to the for loop in function b(). Unfortunately, this is a portion of the for loop that should not be expensive. Adding a fixed amount to an integer is very fast on modern processors, so the samples that oprofile reported were likely attributed to the wrong source line. The line below, test[i]++;, should be very expensive because it accesses the memory subsystem. This line is where the samples should have been attributed.

oprofile can mis-attribute samples for a few reasons beyond its control. First, the processor does not always interrupt on the exact line that caused the event to occur. This may cause samples to be attributed to instructions near the cause of the event, rather than to the exact instruction that caused the event. Second, when source code is compiled, compilers often rearrange instructions to make the executable more efficient. After a compiler has finished optimizing, code may not execute in the order that it was written. Separate source lines may have been rearranged and combined. As a result, a particular instruction may be the result of more than one source line, or may even be a compiler-generated intermediate piece of code that does not exist in the original source. As a result, when the compiler optimizes the code and generates machine instructions, there may no longer be a one-to-one mapping between the original lines of source code and the generated machine instructions. This can make it difficult or impossible for oprofile (and debuggers) to figure out exactly which line of source code corresponds to each machine instruction. However, oprofile tries to be as close as possible, so you can usually look a few lines above and below the line with the high sample count and figure out which code is truly expensive. If necessary, you can use opannotate to show the exact assembly instructions and virtual addresses that are receiving all the samples. It may be possible to figure out what the assembly instructions are doing and then map it back to your original source code by hand. oprofile’s sample attribution is not perfect, but it is usually close enough. Even with these limitations, the profiles provided by oprofile show the approximate source line to investigate, which is usually enough to figure out where the application is slowing down.

4.2.8 Languages: Static (C and C++) Versus Dynamic (Java and Mono)

The majority of the Linux performance tools support analysis of static languages such as C and C++, and all the tools described in this chapter work with applications written in these languages. The tools ltrace, strace, and time work with applications written in dynamic languages such as Java, Mono, Python, or Perl. However, the profiling tools gprof and oprofile cannot be used with these types of applications. Fortunately, most dynamic languages provide non-Linux–specific profiling infrastructures that you can use to generate similar types of profiles.

For Java applications, if the java command is run with the -Xrunhprof command-line option, -Xrunhprof profiles the application. More details are available at http://antprof.sourceforge.net/hprof.html. For Mono applications, if the mono executable is passed the --profile flag, it profiles the application. More details are available at http://www.go-mono.com/performance.html. Perl and Python have similar profile functionality, with Perl’s Devel::DProf described at http://www.perl.com/pub/a/2004/06/25/profiling.html, and Python’s profiler described at http://docs.python.org/lib/profile.html, respectively.