High performance PHP8 at Scale

Running well performing php8 application at scale is challenging itself. We gonna start slow and you gonna listen about Php8 JIT compiler configuration and see couple benchmarks. After the JIT we gonna take to another level. Level of developer good practices. With all that knowledge I will guide you on a path of a greenfield application and show the possible pitfalls and hiccups. Finally we gonna take a ride on a roller-coaster of a legacy app that is performance lacking.

Transcript

1. High performance PHP8 at Scale Max Małecki Poznań, PHPers Summit 27 May 2023 Photo by Oscar Sutton on Unsplash
2. Photo by Drew Stock on Unsplash The Speed
3. Photo by toine G on Unsplash We all love
4. Photo by Jaxon Smith on Unsplash Performance
5. Photo by JOHN BEARBY IMAGES on Unsplash We all want
6. Photo by Bill Jelen on Unsplash Power
7. Source: https://opinion.al/video-sot-36-vjet-nga-vrasja-e-john-lennon/ Is all we need
8. Photo by Isaac Smith on Unsplash Expectations are so...
9. Photo by Dalton Smith on Unsplash High.
10. Photo by Dalton Smith on Unsplash But results may
11. Photo by Racim Amr on Unsplash Vary
12. Photo by David Armstrong on Unsplash Mediocore
13. Photo by Conor Samuel on Unsplash Harmfull Harmfull
14. Greetings!
15. I’m Max I’m from Poznań
16. For 17 years As Senior Backend Engenieer Since 2020
17. Yet another emgiezet/ errbitPHP Talked about technical debt and german cars. Boring speaker bullshit
19. 2023? 2023? I’m Necromancer!
20. Desclimer: Desclimer: I practise what I preach. I practise what I preach. Example will be in Symfony 6 Example will be in Symfony 6 Sorry Tyler no Laravel in this presentation. Sorry Tyler no Laravel in this presentation.
21. Illustrations by Pixeltrue on icons8 0.We gonna talk about “the speed” theory; 1.How to find bottlenecks; 2.PHP Configuration tuning; 3.Every day coding good practices; 4.You will learn about benchmarks methodology; 5.We will watch some cool graphs 6.Apply performance boost in to your project. Today’s Agenda
22. Photo by Drew Stock on Unsplash The Speed
23. Distance traveled in given unit of time
24. v= d t
25. The Speed So how fast your application should respond?
26. The Speed Below 1 second? It is fast enough?
27. The Speed 100 milisecond? It is better?
28. The Speed Latency point of reference
29. L1 cache reference ......................... 0.5 ns Branch mispredict ............................ 5 ns L2 cache reference ........................... 7 ns Mutex lock/unlock ........................... 25 ns Main memory reference ...................... 100 ns Compress 1K bytes with Zippy ............. 3,000 ns = 3 µs Send 2K bytes over 1 Gbps network ....... 20,000 ns = 20 µs SSD random read ........................ 150,000 ns = 150 µs Read 1 MB sequentially from memory ..... 250,000 ns = 250 µs Round trip within same datacenter ...... 500,000 ns = 0.5 ms Read 1 MB sequentially from SSD* ..... 1,000,000 ns = 1 ms Disk seek ........................... 10,000,000 ns = 10 ms Read 1 MB sequentially from disk .... 20,000,000 ns = 20 ms Send packet CA->Netherlands->CA .... 150,000,000 ns = 150 ms from https://gist.github.com/hellerbarde/2843375
30. What the speed really mean? In web request context
31. Requests per second
32. v= r t
33. That gets a response with HTTP_CODE 2**
34. And Meaningfull content. Ofcourse.
35. What the speed really mean? in a consumer / worker context
36. Messages processed per second
37. Jobs finished per second
38. How to measure the speed?
39. Photo by Fotis Fotopoulos on Unsplash Benchmarking
40. Just like a unit tests benchmarks should be organized in suites.
41. Can you give us a example of totally useless benchmark?
43. The Speed PHPBench
44. PHPBench is a benchmark runner for PHP analogous to PHPUnit but for performance rather than correctness.
45. ./vendor/bin/phpbench run tests/Benchmark -- report=default
46. PHPBench (1.2.10) running benchmarks... #standwithukraine with configuration file: /home/mgz/workspace/phpers/summit2023/8.1 Bench/phpbench.json with PHP version 8.0.28, xdebug , opcache ❌ ❌ \emgiezet\Tests\Benchmark\TimeConsumerBench benchConsume............................I4 - Mo152.444μs (±0.09%) Subjects: 1, Assertions: 0, Failures: 0, Errors: 0 +------+-------------------+--------------+-----+------+----------+-----------+--------------+----------------+ | iter | benchmark | subject | set | revs | mem_peak | time_avg | comp_z_value | comp_deviation | +------+-------------------+--------------+-----+------+----------+-----------+--------------+----------------+ | 0 | TimeConsumerBench | benchConsume | | 1000 | 704,944b | 152.652μs | +0.92σ | +0.08% | | 1 | TimeConsumerBench | benchConsume | | 1000 | 704,944b | 152.713μs | +1.38σ | +0.12% | | 2 | TimeConsumerBench | benchConsume | | 1000 | 704,944b | 152.390μs | -1.05σ | -0.09% | | 3 | TimeConsumerBench | benchConsume | | 1000 | 704,944b | 152.390μs | -1.05σ | -0.09% | | 4 | TimeConsumerBench | benchConsume | | 1000 | 704,944b | 152.503μs | -0.20σ | -0.02% | +------+-------------------+--------------+-----+------+----------+-----------+--------------+----------------+
47. With nice customizable HTML reports
49. What is limiting my application speed?
50. Photo by Andrew Seaman on Unsplash Bottlenecks
51. Bottleneck A bottleneck is a phenomenon where the performance and/or capacity of an entire system is limited by a single or limited number of components or resources.
52. Photo by John Cafazza on Unsplash Examples of Examples of
53. Populate 5 million items into ElasticSearch Index.
54. Cronjob that start to overlap the time frame of next execution
55. Elastica update listener on entity change in Doctrine make your insert / updates slow
56. You need to process 15GB XML import overnight
57. You need to serialize big fat entity with lots of relations
58. You want to host mongodb with db bigger than you available RAM
59. Photo by Dylan de Jonge on Unsplash Types of Types of
60. Photo by Dylan de Jonge on Unsplash Hardware Hardware
61. CPU
62. CPU overload
63. CPU Context switches
64. CPU IO waits
65. Memory
66. Out of Memory and go swap
67. Disk
68. Disk: Local disk access latency
69. Disk: Random disk I/O -> disk seeks
70. Disk: Disk fragmentation
71. Disk: SSDs performance drop once near full capacity
72. Photo by Dylan de Jonge on Unsplash Infrastructure Infrastructure
73. Network Long resolving DNS Lookups Dropped packets Faulty hardware
74. Virtualisation Bandwidth to and from the cloud provider Sharing a HDD, disk seek death Network I/O fluctuations in the cloud Enabling high-availability without accounting for failover
75. Database This slide can be separate presentation about database. Most popular: 1)Deadlocks, 2)Large joins taking up memory, 3)Long & short running queries
76. OS Fsync flushing, linux buffer cache filling up, File descriptor limits
77. Photo by Dylan de Jonge on Unsplash Application Application
78. Algorithm complexity
79. Threads
80. Deadlocks
81. No scalability
82. More bottlenecks?
83. How to catch them all?
84. Photo by Andrew Seaman on Unsplash Load Testing Load Testing
85. The goal is to check the performance of your system under a heavy load.
86. To push your system as far as it’s possible and check when it stop responding.
87. Photo: Sekurak.pl Your containers will Your containers will love it. love it.
88. Photo by Lachlan Donald on Unsplash Tools Tools
89. siege siege
90. siege -c 100 -r 100 -b <http://localhost/myapp/>
91. This will run 100 concurent sessions with 100 request each in benchmark mode, so with no delays between requests.
93. { "transactions": 10000, "availability": 100.00, "elapsed_time": 134.82, "data_transferred": 77.36, "response_time": 1.34, "transaction_rate": 74.17, "throughput": 0.57, "concurrency": 99.57, "successful_transactions": 10000, "failed_transactions": 0, "longest_transaction": 1.73, "shortest_transaction": 0.09 }
94. The Speed Apache Benchmark
95. ab -n 100 -c 100 http://localhost/en/blog/posts/aliquam-sodales- odio-id-eleifend-tristique
96. Benchmarking localhost (be patient).....done Server Software: nginx/1.17.8 Server Hostname: localhost Server Port: 80 Document Path: /en/blog/posts/aliquam-sodales-odio-id- eleifend-tristique Document Length: 31309 bytes
97. Concurrency Level: 100 Time taken for tests: 0.271 seconds Complete requests: 100 Failed requests: 0 Total transferred: 3186100 bytes HTML transferred: 3130900 bytes Requests per second: 369.54 [#/sec] (mean) Time per request: 270.604 [ms] (mean) Time per request: 2.706 [ms] (mean, across all concurrent requests) Transfer rate: 11498.08 [Kbytes/sec] received
98. Connection Times (ms) min mean[+/-sd] median max Connect: 0 1 0.2 1 2 Processing: 9 129 73.3 130 259 Waiting: 9 129 73.3 129 259 Total: 10 130 73.1 131 260
99. Percentage of the requests served within a certain time (ms) 50% 131 66% 171 75% 191 80% 205 90% 230 95% 249 98% 260 99% 260 100% 260 (longest request)
100. The Speed Gatling
101. https://github.com/gatling/gatling • Written mostly in Scala • Support load testing scenarios in scala/java/kotlin • Have scenario recorder • Needs own server to unleash its full capabilities • Available also as paid SaaS: gatling.io
102. How to measure changes in my code performance?
103. Put benchmarking as additional CI step
104. Load testing is an overkill to put it in CI.
105. How to find a bottlenecks in my application?
106. Photo by Fotis Fotopoulos on Unsplash Profiling
107. What actually profiling is?
108. Profiling checking the entire callstack for given part of code. Each function is listed with: execution time and amount of calls.
109. XHProf
110. Installation *Example can be found in the presentation repository
111. RUN curl "http://pecl.php.net/get/xhprof-2.3.2.tgz" -fsL -o ./xhprof- 2.3.2.tgz && \ mkdir /var/xhprof && tar xf ./xhprof-2.3.2.tgz -C /var/xhprof && \ cd /var/xhprof/xhprof-2.3.2/extension && \ phpize && \ ./configure && \ make && \ make install # custom settings for xhprof COPY ./xhprof.ini /usr/local/etc/php/conf.d/xhprof.ini RUN docker-php-ext-enable xhprof #folder for xhprof profiles (same as in file xhprof.ini) RUN mkdir -m 777 /profiles Add it to your Dockerfile
112. How to use it
113. #[Route('/', name: 'blog_index', defaults: ['page' => '1', '_format' => 'html'], methods: ['GET'])] #[Route('/rss.xml', name: 'blog_rss', defaults: ['page' => '1', '_format' => 'xml'], methods: ['GET'])] #[Route('/page/{page<[1-9]\d{0,8}>}', name: 'blog_index_paginated', defaults: ['_format' => 'html'], methods: ['GET'])] #[Cache(smaxage: 10)] public function index(Request $request, int $page, string $_format, PostRepository $posts, TagRepository $tags): Response { xhprof_enable(XHPROF_FLAGS_MEMORY + XHPROF_FLAGS_CPU); $tag = null; if ($request->query->has('tag')) { $tag = $tags->findOneBy(['name' => $request->query->get('tag')]); } $latestPosts = $posts->findLatest($page, $tag); file_put_contents('/profiles/'.time().'.application.xhprof', serialize(xhprof_disable())); // Every template name also has two extensions that specify the format and // engine for that template. // See https://symfony.com/doc/current/templates.html#template-naming return $this->render('blog/index.'.$_format.'.twig', [ 'paginator' => $latestPosts, 'tagName' => $tag?->getName(), ]); }
114. Open xhprof web
115. Select the run
116. Select function
117. Pros • Can be used with production settings. • Can profile only couple lines of code.
118. Tip of the day. • If you like your SSD life span turn off profiling while load testing ;)
119. Other options?
120. xdebug
121. Symfony Profiler
124. How to improve app response speed?
125. Photo by Fotis Fotopoulos on Unsplash Caching
126. How the load will look like when app hasn’t any cache?
128. 2 Goals:
129. 1. Response cache: should prevent duplicated requests to reach the php-fpm.
130. 2. Result cache: should prevent cached queries to reach the database.
131. If the cache entry has valid TTL.
132. Cache keys should have semantic naming
133. Because
134. There are only two hard things in Computer Science: cache invalidation and naming things. - Phil Karlton
135. Check your headers! Cache-Control: max-age=604800, must-revalidate
136. REST API GET endpoint responses can be cached
137. Database is last thing that should be called.
138. Prioritize endpoints caching by load they generate.
139. Photo by Kelly Sikkema on Unsplash Intro done Now we’re on the same page
140. Photo by Kyle Mackie on Unsplash It’s time for the main course
141. Photo by shri on Unsplash or for the tofu course To don’t offend anybody today.
142. PHP Configuration
143. memory_limit=1024M opcache.enable=1 opcache.revalidate_freq=10 opcache.validate_timestamps=1 opcache.max_accelerated_files=10000 opcache.memory_consumption=192 opcache.max_wasted_percentage=10 opcache.interned_strings_buffer=1 opcache.fast_shutdown=1 opcache.jit_buffer_size=256M opcache.jit=1255 opcache.jit=tracing upload_max_filesize = 20M post_max_size = 20M pm = ondemand pm.max_children = 500 pm.max_requests = 500
144. Composer
146. /usr/bin/composer install \ --no-progress \ --no-interaction \ --no-dev \ --optimize-autoloader \ --prefer-dist
147. composer dump-autoload --classmap- authoritative
148. Or put this to APCu
149. composer dump-autoload --apcu
150. OPCache
151. Installation apt-get install php-opcache
152. Configuration #/etc/php8.1/fpm/php.ini opcache.enable=1 opcache.memory_consumption=128 opcache.max_accelerated_files=10000 opcache.revalidate_freq=200
153. Restart php-fpm
154. Add some extra preloading on fpm startup composer require practically/preloader
155. practically/preloader hooks after install and generates preload.php executable file.
156. update php.ini ;/etc/php8.1/fpm/php.ini opcache.enable=1 opcache.enable_cli=1 opcache.memory_consumption=128 opcache.max_accelerated_files=10000 opcache.revalidate_freq=200 opcache.preload=/path/to/preload.php opcache.preload_user=user
157. JIT Compiler
158. opcache.jit_buffer_size=256M opcache.jit=1255 opcache.jit=tracing
159. JIT Modes • opcache.jit = 1205 - all code is JIT compiled • opcache.jit = 1235 - only selected code portions (based on their relative use) are passed to the JIT compilation • opcache.jit = 1255 - application code is tracked for compilation by JIT and selected parts of the code are transferred to the compiler
160. Photo by Fotis Fotopoulos on Unsplash Every day coding good practices
161. Use isset() to check that array is initialized
162. \emgiezet\Tests\Benchmark\IssetBench benchIsset..............................I0 - Mo0.012μs (±0.00%) benchSizeof.............................I0 - Mo0.021μs (±0.00%) benchCount..............................I0 - Mo0.021μs (±0.00%)
163. Stick with single quotes
164. \emgiezet\Tests\Benchmark\QuotesBench benchSingle.............................I0 - Mo0.445μs (±0.00%) benchDouble.............................I0 - Mo2.498μs (±0.00%)
165. RegExp’s are sexy but you need to pay
166. \emgiezet\Tests\Benchmark\PregBench benchPregReplace........................I0 - Mo3.906μs (±0.00%) benchStrReplace.........................I0 - Mo2.832μs (±0.00%)
167. Try identical operator instead of equals
168. \emgiezet\Tests\Benchmark\EqualsBench benchEqual..............................I0 - Mo0.305μs (±0.00%) benchIdentical..........................I0 - Mo0.301μs (±0.00%)
169. When you need precision math use BCMath but be aware that it costs double
170. \emgiezet\Tests\Benchmark\MathBench benchBcmul..............................I0 - Mo0.492μs (±0.00%) benchMultiply...........................I0 - Mo0.258μs (±0.00%)
171. Forget about other loops than foreach()
172. \emgiezet\Tests\Benchmark\LoopsBench benchFor................................I0 - Mo7.168ms (±0.00%) benchForeach............................I0 - Mo6.791ms (±0.00%) benchWhile..............................I0 - Mo7.153ms (±0.00%) benchDoWhile............................I0 - Mo7.056ms (±0.00%)
173. If you want more benchmarks like those go: https://phpbench.com/
174. If you want to run my benchmarks yourself go: https://github.com/emgiezet/ phpers-summit-2023
175. Find balance between Lazy & Eager loading in Doctrine
176. Use Generators & Iterators
177. Trust me. Check @alexdaubois presentation at 13:45
178. Remember of batch processing
179. Partition your cli commands
180. Never try to load XML ’s bigger than 50 MB to DOMObject
181. Don’t spam logs. Each echo() each print() costs.
182. Photo by Fotis Fotopoulos on Unsplash Benchmark time!
183. Benchmark App • Symfony Demo App – In a docker container – Ngnix & php-fpm – In prod mode – Whitout xdebug – Mysql isntead of sqlite – It has symfony.cache enabled (#[Cache(smaxage: 10)])
185. What we gonna benchmark? • 8.1 • 8.1 Opcache • 8.1 Opcache + JIT compiler enabled
186. How we gonna benchmark? • Siege on multiple urls siege -c 100 -r 100 -b -i -f urls.txt
187. PHP 8.1 – no opcache { "transactions": 66730, "availability": 100.00, "elapsed_time": 149.54, "data_transferred": 5511.54, "response_time": 0.22, "transaction_rate": 446.24, "throughput": 36.86, "concurrency": 99.45, "successful_transactions": 66730, "failed_transactions": 0, "longest_transaction": 1.75, "shortest_transaction": 0.00 }
188. PHP 8.1 – opcache { "transactions": 66988, "availability": 100.00, "elapsed_time": 24.26, "data_transferred": 5535.86, "response_time": 0.04, "transaction_rate": 2761.25, "throughput": 228.19, "concurrency": 99.38, "successful_transactions": 66988, "failed_transactions": 0, "longest_transaction": 0.40, "shortest_transaction": 0.00 }
189. PHP 8.1 – opcache & jit { "transactions": 66724, "availability": 100.00, "elapsed_time": 19.69, "data_transferred": 5510.57, "response_time": 0.03, "transaction_rate": 3388.73, "throughput": 279.87, "concurrency": 99.36, "successful_transactions": 66724, "failed_transactions": 0, "longest_transaction": 0.45, "shortest_transaction": 0.00 }
190. Req/s boost Php 8.1 Opcache Enabled Opcache and JIT Enabled 0 500 1000 1500 2000 2500 3000 3500 4000 PHP 8.1 Performance Request / s
191. Photo by Brands&People on Unsplash What you can What you can apply to your apply to your project? project?
193. Photo by Azin Javadzadeh on Unsplash Greenfield
194. Manage the performance expectations
195. Run loadtests after each delivery cycle.
196. Add benchmarks to your CI pipeline.
197. Try to align your benchmark scenarios with KPI that businesses wants.
198. Take baby steps
199. Don’t over-engineer
201. Add static code analysis. With performance rules check.
202. Always consider performance impact on feature planning.
203. Always consider performance impact on code review.
204. First implementation is always rubbish.
205. Photo by Brett Jordan on Unsplash Legacy Legacy
206. Set up performance goals.
207. Find bottlenecks.
208. Prioritize them with businesses.
209. Decisions should be supported by real data.
210. Plan them. Apply one change at the time and then benchmark!
211. And… Repeat from the first step.
213. Photo by Fotis Fotopoulos on Unsplash Any Questions?
214. Catch me! • Code from presentation: https://github.com/emgiezet/phpers-summit- 2023 • Twitter: @mgz
215. Photo by Kelly Sikkema on Unsplash