Click here for a live demo of the analysis screen
SPX, which stands for Simple Profiling eXtension, is just another profiling extension for PHP. It differentiates itself from other similar extensions as being:
- totally free and confined to your infrastructure (i.e. no data leaks to a SaaS).
- very simple to use: just set an environment variable (command line) or switch on a radio button (web page) to profile your script. Thus, you are free of:
- manually instrumenting your code (Ctrl-C a long running command line script is even supported).
- using a dedicated browser extension or command line launcher.
- multi metrics capable: 21 currently supported (various time metrics, memory, included files, objects in use, I/O...).
- able to collect data without losing context. For example Xhprof (and potentially its forks) aggregates data per caller / callee pairs, which implies the loss of the full call stack and forbids timeline or Flamegraph based analysis.
- shipped with its web UI which allows to:
- enable / configure profiling for the current browser session
- list profiled script reports
- select a report for in-depth analysis, featuring these interactive visualizations:
- timeline (scale to millions of function calls)
- flat profile
- Flamegraph
Platforms support is currently quite limited. Feel free to open an issue if your platform is not supported. Current requirements are:
- x86-64
- GNU/Linux or macOS
- zlib dev package (e.g. zlib1g-dev on Debian based distros)
- PHP 5.6 & 7+
- Non-ZTS (threaded) build of PHP (ZTS support is theoretical)
- PHP development package (corresponding to your installed PHP version).
- zlib development package:
- For Debian based distros (including Ubuntu, Kubuntu...), just run:
sudo apt-get install zlib1g-dev.
- For Debian based distros (including Ubuntu, Kubuntu...), just run:
git clone https://github.com/NoiseByNorthwest/php-spx.git
cd php-spx
phpize
./configure
make
sudo make installThen add extension=spx.so to your php.ini, or in a dedicated spx.ini file created within the include directory.
You may also want to override default SPX configuration to be able to profile a web page, with this one for example for a local development environment.
On GNU/Linux, SPX uses procfs (i.e. by reading files under /proc directory) to get I/O stats of the current thread. This is what is done under the hood when you select at least one of these 3 metrics: io, ior or iow.
But, on most PHP-FPM setups, you will have a permission issue preventing SPX to open /proc/self/task/<thread Id>/io.
This is due to the fact that PHP-FPM master process runs as root when child processes run as another unprivileged user.
When this is the case, the process.dumpable = yes line must be added to the FPM pool configuration so that child processes will be able to read /proc/self/task/<thread Id>/io.
This is still experimental. API might change, features might be added or dropped, or development could be frozen.
You can still safely use it in a non-production environment.
Contributions are welcome but be aware of the experimental status of this project and please follow the contribution rules described here: CONTRIBUTING.md
Assuming a development environment with the configuration described here and your application is accessible via http://localhost.
Just open with your browser the following URL: http://localhost/?SPX_KEY=dev&SPX_UI_URI=/ to access to the web UI control panel.
N.B.: I assume http://localhost/ is served by a PHP script through standard web server feature like directory index or URL rewriting. If this is not the case you can also explicitly point to an existing PHP script as in the following example: http://localhost/my_application.php?SPX_KEY=dev&SPX_UI_URI=/. The PHP script will not be executed, SPX will intercept and disable its execution to serve its content in place.
You will then see the following form:
Then switch on "Enabled". At this point profiling is enabled for the current domain and your current browser session through a set of dedicated cookies.
Then refresh the web page you want to profile and refresh the control panel to see the generated report in the list below the control panel form.
Then click on the report in the list and enjoy the analysis screen.
Just prepend your command line with SPX_ENABLED=1 to trigger profiling. You will get the flat profile printed on STDERR at the end of the execution, even if you abort it by hitting Ctrl-C, as in the following example:
$ SPX_ENABLED=1 composer update
Loading composer repositories with package information
Updating dependencies (including require-dev)
^C
*** SPX Report ***
Global stats:
Called functions : 27.5K
Distinct functions : 714
Wall time : 7.39s
ZE memory : 62.6MB
Flat profile:
Wall time | ZE memory |
Inc. | *Exc. | Inc. | Exc. | Called | Function
----------+----------+----------+----------+----------+----------
101.6ms | 101.6ms | 41.8MB | 41.8MB | 12 | Composer\Json\JsonFile::parseJson
53.6ms | 53.6ms | 544B | 544B | 4 | Composer\Cache::sha256
6.91s | 41.5ms | 41.5MB | -7.5MB | 4 | Composer\Repository\ComposerRepository::fetchFile
6.85s | 32.3ms | 47.5MB | 5.4MB | 5 | 1@Composer\Repository\ComposerRepository::loadProviderListings
7.8ms | 7.8ms | 0B | 0B | 4 | Composer\Cache::write
1.1ms | 1.1ms | -72B | -72B | 1 | Composer\Console\Application::Composer\Console\{closure}
828.5us | 828.5us | 976B | 976B | 12 | Composer\Util\RemoteFilesystem::findHeaderValue
497.6us | 491.0us | 710.2KB | 710.2KB | 1 | Composer\Cache::read
2.4ms | 332.6us | 20.9KB | -378.8KB | 34 | 3@Symfony\Component\Finder\Iterator\FilterIterator::rewind
298.9us | 298.9us | 2.2KB | 2.2KB | 47 | Symfony\Component\Finder\Iterator\FileTypeFilterIterator::acceptN.B.: Just add SPX_FP_LIVE=1 to enable the live refresh of the flat profile during script execution.
You just have to specify SPX_REPORT=full to generate a report available via the web UI:
SPX_ENABLED=1 SPX_REPORT=full ./bin/console cache:clearIf your CLI script is long-living and/or daemonized (e.g. via supervisord), profiling its whole lifespan could be meaningless. This is especially true in case of a service waiting for tasks to process.
To handle this case, SPX allows to disable the automatic start of profiling and exposes 2 userland functions, spx_profiler_start(): void & spx_profiler_stop(): void, in order to respectively control the start and the end of the profiled spans.
Here is how you can instrument your script:
<?php
while ($task = get_next_ready_task()) {
spx_profiler_start();
try {
$task->process();
} finally {
spx_profiler_stop();
}
}And of course this script must be run at least with profiling enabled and the automatic start disabled as in the following command:
SPX_ENABLED=1 SPX_REPORT=full SPX_AUTO_START=0 my_script.phpSide notes:
spx_profiler_start()andspx_profiler_stop()can safely be nested.- when automatic start is disabled, no signal handlers (i.e. on SIGINT/SIGTERM) are registered by SPX.
- automatic start can only be disabled for CLI SAPI.
| Name | Default | Changeable | Description |
|---|---|---|---|
| spx.data_dir | /tmp/spx |
PHP_INI_SYSTEM | The directory where profiling reports will be stored. You may change it to point to a shared file system for example in case of multi-server architecture. |
| spx.http_enabled | 0 |
PHP_INI_SYSTEM | Whether to enable web UI and HTTP request profiling. |
| spx.http_key | PHP_INI_SYSTEM | The secret key used for authentication (see security concern for more details). You can use the following command to generate a 16 bytes random key as an hex string: openssl rand -hex 16. |
|
| spx.http_ip_var | REMOTE_ADDR |
PHP_INI_SYSTEM | The $_SERVER key holding the client IP address used for authentication (see security concern for more details). Overriding the default value is required when your application is behind a reverse proxy. |
| spx.http_trusted_proxies | 127.0.0.1 |
PHP_INI_SYSTEM | The trusted proxy list as a comma separated list of IP addresses. This setting is ignored when spx.http_ip_var's value is REMOTE_ADDR. |
| spx.http_ip_whitelist | PHP_INI_SYSTEM | The IP address white list used for authentication as a comma separated list of IP addresses, use * to allow all IP addresses. |
|
| spx.http_ui_assets_dir | /usr/local/share/misc/php-spx/assets/web-ui |
PHP_INI_SYSTEM | The directory where the web UI files are installed. In most cases you do not have to change it. |
For your local & private development environment, since there is no need for authentication, you can use this configuration:
spx.http_enabled=1
spx.http_key="dev"
spx.http_ip_whitelist="127.0.0.1"
And then access to the web UI at http(s)://<your application host>/?SPX_KEY=dev&SPX_UI_URI=/.
Here is the list of available metrics to collect. By default only Wall time and Zend Engine memory usage are collected.
| Key (command line) | Name | Description |
|---|---|---|
| wt | Wall time | The absolute elapsed time. |
| ct | CPU time | The time spent while running on CPU. |
| it | Idle time | The time spent off-CPU, that means waiting for CPU, I/O completion, a lock acquisition... or explicitly sleeping. |
| zm | Zend Engine memory usage | Equivalent to memory_get_usage(false). |
| zmac | Zend Engine allocation count | Number of memory allocations (i.e. allocated blocks) performed. |
| zmab | Zend Engine allocated bytes* | Number of allocated bytes. |
| zmfc | Zend Engine free count | Number of memory releases (i.e. freed blocks) performed. |
| zmfb | Zend Engine freed bytes* | Number of freed bytes. |
| zgr | Zend Engine GC run count | Number of times the GC (cycle collector) have been triggered (either manually or automatically). |
| zgb | Zend Engine GC root buffer length | Root buffer length, see explanation here. It could be helpful to track pressure on garbage collector. |
| zgc | Zend Engine GC collected cycle count | Total number of collected cycles through all GC runs. |
| zif | Zend Engine included file count | Number of included files. |
| zil | Zend Engine included line count | Number of included lines. |
| zuc | Zend Engine user class count | Number of userland classes. |
| zuf | Zend Engine user function count | Number of userland functions (including userland class/instance methods). |
| zuo | Zend Engine user opcode count | Number of included userland opcodes (sum of all userland file/function/method opcodes). |
| zo | Zend Engine object count | Number of objects currently held by user code. |
| ze | Zend Engine error count | Number of raised PHP errors. |
| io | I/O (reads + writes)** | Bytes read or written while performing I/O. |
| ior | I/O (reads)** | Bytes read while performing I/O. |
| iow | I/O (writes)** | Bytes written while performing I/O. |
*: Allocated and freed byte counts will not be collected if you use a custom allocator or if you force the libc one through the USE_ZEND_ALLOC environment variable set to 0.
**: I/O metrics are not supported on macOS. On GNU/Linux you should read this if you use PHP-FPM.
Contrary to web page profiling which only support full report type (the one exploitable by the web UI), command line script profiling supports several types of report. Here is the list below:
| Key | Name | Description |
|---|---|---|
| fp | Flat profile | The flat profile provided by SPX. It is the default report type and is directly printed on STDERR. |
| full | Full report | This is the report type for web UI. Reports will be stored in SPX data directory and thus will be available for analysis on web UI side. |
| trace | Trace file | A custom format (human readable text) trace file. |
| Name | Default | Description |
|---|---|---|
| SPX_ENABLED | 0 |
Whether to enable SPX profiler (i.e. triggering profiling). When disabled there is no performance impact on your application. |
| SPX_AUTO_START | 1 |
Whether to enable SPX profiler's automatic start. When automatic start is disabled, you have to start & stop profiling on your own at runtime via the spx_profiler_start() & spx_profiler_stop() functions. See here for more details. |
| SPX_BUILTINS | 0 |
Whether to profile internal functions, script compilations, GC runs and request shutdown. |
| SPX_DEPTH | 0 |
The stack depth at which profiling must stop (i.e. aggregate measures of deeper calls). 0 (default value) means unlimited. |
| SPX_SAMPLING_PERIOD | 0 |
Whether to collect data for the current call stack at regular intervals according to the specified sampling period (0 means no sampling). The result will usually be less accurate but in some cases it could be far more accurate by not over-evaluating small functions called many times. It is recommended to try sampling (with different periods) if you want to accurately find a time bottleneck. When profiling a long running & CPU intensive script, this option will allow you to contain report size and thus keeping it small enough to be exploitable by the web UI. See here for more details. |
| SPX_METRICS | wt,zm |
Comma separated list of available metric keys to collect. All report types take advantage of multi-metric profiling. |
| SPX_REPORT | fp |
Selected report key. |
| SPX_FP_FOCUS | wt |
Metric key for flat profile sort. |
| SPX_FP_INC | 0 |
Whether to sort functions by inclusive value instead of exclusive value in flat profile. |
| SPX_FP_REL | 0 |
Whether to display metric values as relative (i.e. percentage) in flat profile. |
| SPX_FP_LIMIT | 10 |
The flat profile size (i.e. top N shown functions). |
| SPX_FP_LIVE | 0 |
Whether to enable flat profile live refresh. Since it plays with cursor position through ANSI escape sequences, it uses STDOUT as output, replacing script output (both STDOUT & STDERR). |
| SPX_FP_COLOR | 1 |
Whether to enable flat profile color mode. |
| SPX_TRACE_SAFE | 0 |
The trace file is by default written in a way to enforce accuracy, but in case of process crash (e.g. segfault) some logs could be lost. If you want to enforce durability (e.g. to find the last event before a crash) you just have to set this parameter to 1. |
| SPX_TRACE_FILE | Custom trace file name. If not specified it will be generated in /tmp and displayed on STDERR at the end of the script. |
Well, as you might already noticed in corresponding basic usage example, setting a SPX parameter for a command line script simply means setting an environment variable with the same name.
Since the web UI uses advanced JavaScript features, only the following browsers are known to be supported:
- most recent version of any Chromium-based browser.
- most recent version of Firefox with
dom.moduleScripts.enabledpreference set totrue.
This is the home page of the web UI, divided into 2 parts:
- the control panel for setting the profiling setup for your current browser session.
- the profile report list as a sortable table. A click on a row allows to go to the analysis screen for the corresponding report.
Click here for a live demo of the analysis screen
The analysis screen can nicely handle profile reports with up to several (5+) millions of recorded function calls with Chromium on my i5 @ 3.3GHz / 8GB desktop. In case you want to profile a long running, CPU intensive, script which tends to generate giant reports, you can enable sampling mode with the suitable sampling period. See SPX_SAMPLING_PERIOD parameter for command line script.
This is simply a combo box for selecting the currently analyzed metric.
By default, function related blocks in the visualizations are colored according to their cost, with a color scale displayed at the top right of the screen.
You can also define a custom color scheme by clicking on the color scheme mode link, displayed at the top of the screen just after the metric selector.
A drop-down window will then appear and allow you to switch between default and category mode and define (add/edit/delete) your categories (color, name, pattern list) for the category mode (see the screenshot below).
This visualization is the timeline overview of all called functions. You can change the selected time range by, represented by a transparent green rectangle, by simply dragging it horizontally.
Except for wall time, the current metric is also plotted (current value over time) on a foreground layer.
Supported controls:
- horizontal left click drag: shift the selected time range
- resize click on selected time range rectangle: shift one of the selected time range boundary
This visualization is an interactive timeline which is able to control and keep focus on the selected time range.
Supported controls:
- left click drag: time range shift (horizontal) or depth range shift (vertical)
- middle click vertical drag: time range zoom in/out
- mouse wheel: time range zoom in/out
- hovering a function call to show more details
- double click on a function call: set the current time range as the one of the selected function call
Except for wall time, the current metric is also plotted (current value over time) on a foreground layer.
This visualization is the flat profile for the selected time range and the selected metric, displayed as a sortable table.
This visualization, designed by Brendan Gregg, allows to quickly find the hot code path for the selected time range and the selected metric. Metrics corresponding to releasable resources (memory, objects in use...) are not supported by this visualization.
You can highlight a function by clicking on one of its spans within the timeline or Flamegraph widgets or its name within the flat profile widget.
The lack of review / feedback about this concern is the main reason SPX cannot yet be considered as production ready.
SPX allows you to profile web page as well as command line scripts, and also to list and analyze profile reports through its embedded web UI. This is why there is a huge security risk, since an attacker could:
- access to web UI and get sensible information about your application.
- to a lesser extent, make a DoS attack against your application with a costly profiling setup.
So, unless access to your application is already restricted at lower layer (i.e. before your application is hit, not by the application / PHP framework itself), a client triggering profiling or accessing to the web UI must be authenticated.
SPX provides two-factor authentication with these 2 mandatory locks:
- IP address white list (exact string representation matching).
- Fixed secret random key (generated on your own) provided via a request header, cookie or query string parameter.
Thus a client can profile your application via a web page only if its IP address is white listed and its provided key is valid.
In tracing mode (default), SPX is subject to accuracy issues for time related metrics when the measured function execution time is:
- close or lower than the timer precision
- close or lower than SPX's own per function overhead
The first issue is mitigated by using the highest resolution timer provided by the platform. On Linux & recent macOS versions the timer resolution is 1ns; on macOS before 10.12/Sierra, the timer resolution is only 1us.
The second issue is mitigated by taking into account SPX's time (wall / cpu) overhead by subtracting it to measured function execution time. This is done by evaluating SPX constant per function overhead before starting profiling the script.
However, whatever the platform, if you want to maximize accuracy to find a time bottleneck, you should also:
- avoid profiling internal functions.
- avoid collecting additional metrics.
- try sampling mode with different sampling periods.
- try to play with maximum depth parameter to stop profiling at a given depth.
I have found lot of inspiration and hints reading: