optimizing PHP-FPM, Nginx and kernel for high performance

PHPIstanbul 2016
Video https://www.youtube.com/watch?v=9BDdREjmD5A

Transcript

1. optimizing PHP-FPM, Nginx and kernel for high performance osman ungur, github.com/o
2. osman ungur software (still i can't find a good name yet) developer with sysadmin background freelance, consultant, site reliability engineer avid runner, cyclist
3. java, php, a little bit python and erlang spring, hibernate, symfony, doctrine big data, nosql tools, scalability optimization, automation, monitoring
4. today we will see problems with selling lemonade problems with running an app on server why tuning and capacity planning is important key configuration parameters
5. imagine, you're selling lemonade for profit
6. you need to buy • lemons • ice cubes • cups • sugars
7. make sure you found the best recipe good location
8. mostly, you can expect how many cups of lemonade you can sell
9. • weather • advertisement • news • location • satisfaction • popularity
10. limited money limited stocks / supplies limited storage
11. good resource planning before advertisements hot weather conditions
12. out of stock problems out of (lemon|ice|cup|sugar) loss of popularity loss of money
13. over-stock problems spoiled lemons melting ices waste of resources loss of money
14. you need to find a sweet spot
15. most important things reduced wait times customer satisfaction great recipe
16. key factors economics resource management capacity planning
17. imagine, you built an app
18. you need a server virtual machine container
19. make sure you found the best servers good location
20. mostly, you can expect how many visitors wants to use your app
21. • advertisement • news • slashdot-effects • satisfaction • popularity
22. limited money limited cpu / memory limited storage / io
23. good resource planning before advertisements hot sales
24. out of resources haproxy[12947]: backend app_api_backend has no server available! nginx[432]: worker_connections exceed open file resource limit: 1024 [crit]: connect() to unix:/tmp/php5-fpm.sock failed
25. more resources ineffecient cpu usage waste of memory loss of money
26. you need to find a sweet spot
27. most important things reduced wait times customer satisfaction great configuration recipe
28. key factors economics resource management capacity planning
29. monitoring using a load-balancer alerts, dashboards a/b tests for different configurations
30. make your applications horizontally scalable avoid single point of failure easy to provision
31. avoid using default configurations
32. fail fast
33. optimizating nginx
34. worker_process is the backbone of nginx workers are not multi-threaded so they do not spread the per-connection across CPU cores usually 1 worker per CPU core
35. worker_connections effectively limits how many connections each worker can maintain at a time if you grow in traffic you’ll want to eventually increase the amount of connections each worker can do
36. keep_alive is a HTTP feature which allows user agents to keep the connection keep alive connections have such a huge impact on nginx performance You probably don’t need a keep alive time out value of 65 but 10-20 is definitely recommended
37. enable tcp_nodelay and tcp_nopush these directives determine how the OS handles the network buffers and when to flush them to the end user disable if you are running in vagrant for dev purposes
38. minimize i/o disable access logs send it over syslog or use logging capability of loadbalancer less i/o and easy to spot problems only enable `warn` level errors in logs
39. use open_file_cache caches open file descriptors, their sizes and modification times information on existence of directories file lookup errors
40. adjust buffers If the buffer sizes are set too low nginx will have to store the responses from upstreams in a temporary file which causes both write and read IO fastcgi_buffer_size, fastcgi_buffers, fastcgi_*_timeout
41. reduce network io use gzip find a good gzip_comp_level and gzip_min_length flush once
42. use fastcgi_cache for storing and serving php-fpm responses only for `Public` and non-user specific 200 responses also cache 301's and 404's "Here be dragons"
43. php-fpm optimization
44. avoid unix socket limited scalability %5-%10 faster than TCP listen = 127.0.0.1:9000
45. minimize listen backlog or disable if possible if you need backlog, you're doing wrong listen.backlog = 0
46. php-fpm controls child processes based on different strategies and parameters you need to be aware how many processes used and waiting for sending responses
47. `static` process manager consistent memory usage minimizes context switch simple pm.max_children = 20
48. `dynamic` process manager inconsistent memory and cpu usage child processes are set dynamically based on the parameters pm.start_servers, pm.*_spare_servers default values not tuned for production purposes
49. `ondemand` process manager no children are created at startup children will be forked when new requests will connect very inconsistent memory and cpu usage can hogs your server quickly
50. `pm.max_requests` number of requests each child process should execute before respawning limiting this avoids memory leaks you should use this
51. use `pm.status_path` and proxy over nginx you can monitor a lot of important metric from this endpoint also you can see how many workers is busy and plan for changing worker process parameters
52. fail fast! use request_terminate_timeout also set max_execution_time in php.ini avoid infinite responses find a reasonable number
53. use and optimize zend_opcache find good numbers for max_accelerated_files memory_consumption and interned_strings_buffer disable revalidation in production
54. kernel optimization
55. minimize or disable swappiness reduce unnecessary io you can set from sysctl eg: vm.swappiness: 10
56. increase ipv4.ip_local_port_range because nginx and php-fpm uses tcp in request lifecycle using php-fpm doubles port usage
57. tune net.ipv4* and net.core* values in sysctl this parameters determines how OS handles buffers, time-waits, backlogs and timeouts use carefully!
58. tune open_file_limit from sysctl and /etc/security/limits.conf ensure openfiles and process limits configured for user running of php-fpm and nginx
59. i don't have magic configurations all parameters based on your workload application characteristics and server configuration monitoring helps you to find good numbers
60. monitoring for all the things!
61. ignorance is bliss - Thomas Gray
62. wrapping things up problems with selling lemonade problems with running an app on server why tuning and capacity planning is important key configuration parameters
63. questions? thanks! github.com/o osman ungur