Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

No Callbacks, No Threads - RailsConf 2010


Published on

Multi-threaded servers compete for the global interpreter lock (GIL) and incur the cost of continuous context switching, potential deadlocks, or plain wasted cycles. Asynchronous servers, on the other hand, create a mess of callbacks and errbacks, complicating the code. But, what if, you could get all the benefits of asynchronous programming, while preserving the synchronous look and feel of the code – no threads, no callbacks?

Published in: Technology

No Callbacks, No Threads - RailsConf 2010

  1. No Callbacks, No Threads & Ruby 1.9<br />async & co-operative web-servers<br />Ilya Grigorik<br />@igrigorik<br />
  2. The slides…<br />Twitter<br />My blog<br />
  3. The state of art is not good enough.<br />(we’ve been stuck in the same local minima for several years)<br />
  4. require"active_record”<br />ActiveRecord::Base.establish_connection(<br /> :adapter => "mysql",<br /> :username => "root",<br /> :database => "database",<br /> :pool => 5<br />)<br />threads = []<br />10.times do |n| <br /> threads << {<br />ActiveRecord::Base.connection_pool.with_connectiondo |conn|<br /> res =conn.execute("select sleep(1)")<br />end<br /> }<br />end<br />threads.each { |t| t.join }<br />The “Experiment”<br />vanilla everything…<br />
  5. require"active_record”<br />ActiveRecord::Base.establish_connection(<br /> :adapter => "mysql",<br /> :username => "root",<br /> :database => "database",<br /> :pool => 5<br />)<br />threads = []<br />10.times do |n| <br /> threads << {<br />ActiveRecord::Base.connection_pool.with_connectiondo |conn|<br /> res =conn.execute("select sleep(1)")<br />end<br /> }<br />end<br />threads.each { |t| t.join }<br />5 shared connections<br /># time ruby activerecord-pool.rb<br />#<br /># real 0m10.663s<br /># user 0m0.405s<br /># sys 0m0.201s<br />
  6. BHP<br />% power <br /> loss<br />WHP<br />
  7. > 50% power loss!?<br />
  8. Mongo<br />Couch<br />MySQL<br />PSQL<br />…<br />Drivers<br />Threads<br />Ruby VM<br />GIL<br />Fibers<br />…<br />Network<br />Mongrel<br />Unicorn<br />Passenger<br />…<br />
  9. 2<br />Mongo<br />Couch<br />MySQL<br />PSQL<br />…<br />Drivers<br />Threads<br />Ruby VM<br />1<br />4<br />GIL<br />Fibers<br />…<br />Network<br />Mongrel<br />Unicorn<br />3<br />We’re as fast as the slowest component<br />Passenger<br />…<br />
  10. Global Interpreter Lock is a mutual exclusion lock held by a programming language interpreter thread to avoid sharing code that is not thread-safe with other threads. <br />There is always one GIL for one interpreter process.<br />Concurrency is a myth in Ruby<br />(with a few caveats, of course)<br /><br />
  11. N-M thread pool in Ruby 1.9…<br />Better but still the same problem!<br />Concurrency is a myth in Ruby<br />still no concurrency in Ruby 1.9<br /><br />
  12. Nick – tomorrow @ 11:45am<br />Concurrency is a myth in Ruby<br />still no concurrency in Ruby 1.9<br /><br />
  13. Blocks entire<br />Ruby VM<br />Not as bad, but<br />avoid it still..<br />Avoid locking interpreter threads at all costs<br />let’s say you’re writing an extension…<br />
  14. Will fetch_xyz() block the VM?<br />when was the last time you asked yourself this question?<br />
  15. require 'rubygems’<br />require 'sequel'DB = Sequel.connect('mysql://root@localhost/test')while trueDB['select sleep(1)'].select.firstend<br />Blocking 1s call!<br />ltrace –ttTg -xmysql_real_query –p ./example.rb<br />mysql.gem under the hood<br />22:10:00.218438 mysql_real_query(0x02740000, "select sleep(1)", 15) = 0 <1.001100>22:10:01.241679 mysql_real_query(0x02740000, "select sleep(1)", 15) = 0 <1.000812><br /><br />
  16. Blocking calls to mysql_real_query<br />mysql_real_query requires an OS thread<br />Blocking on mysql_real_query blocks the Ruby VM<br />Aka, “select sleep(1)” blocks the entire Ruby runtime for 1s<br />(ouch)<br />gem install mysqlwhat you didn’t know…<br />
  17. static VALUE async_query(intargc, VALUE* argv, VALUE obj) {<br /> ...<br />send_query( obj, sql );<br /> ...<br />schedule_query( obj, timeout);<br /> ...<br />returnget_result(obj); <br />}<br />staticvoidschedule_query(VALUEobj, VALUE timeout) {<br /> ...<br />structtimevaltv = { tv_sec: timeout, tv_usec: 0 };<br />for(;;){<br />FD_ZERO(&read);<br />FD_SET(m->net.fd, &read);<br /> ret = rb_thread_select(m->net.fd + 1, &read, NULL, NULL, &tv);<br /> ...<br />if (m->status == MYSQL_STATUS_READY)<br />break;<br /> }<br />}<br />send query and block<br />Ruby: select() = C: rb_thread_select()<br />mysqlplus.gem under the hood<br />
  18. spinning in select<br />mysqlplus.gem + ruby select<br />
  19. Step 1: Fix the drivers<br />Many of our DB drivers don’t respect the <br />underlying Ruby VM. Don’t blame the VM.<br />
  20. Drivers<br />Ruby VM<br />Network<br />WEBRick<br />Mongrel made Rails viable<br />still powers a lot of Rails apps today<br />
  21. Rails rediscovers Apache<br />the worker/forker model… <br />
  22. *nix IPC is fast! Woo!<br />…<br />Full Ruby VM<br />An exclusive Ruby VM for EACH request<br />am I the only one who thinks this is terrible? <br />
  23. “Does not care if your application is thread-safe or not, workers all run within their own isolated address space and only serve one client at a time for maximum robustness.”<br />Robustness? That sounds like a bug.<br />An exclusive Ruby VM for EACH request<br />am I the only one who thinks this is terrible? <br />
  24. Step 2: consider entire stack<br />The driver, the web-server, and the network<br />must all work together.<br />
  25. Node imposes the full-stack requirements<br />Node imposes async drivers<br />Node imposes async frameworks<br />Surprise: Node is “fast”<br />
  26. We can ignore the performance<br />issues at our own peril<br />or, we can just fix the problem<br />
  27. ><br />I’ll take Ruby over JS<br />gem install eventmachine<br />
  28. p "Starting"EM.rundop"Running in EM reactor"endp”won’t get here"<br />whiletruedo<br /> timersnetwork_ioother_io<br />end<br />EventMachine Reactor<br />concurrency without thread<br />EventMachine: The Speed DemonWednesday @ 11:45am – Aman Gupta<br />
  29. Non-blocking IO requires non-blocking drivers:<br />AMQP<br />MySQLPlus<br />Memcached<br />DNS<br />Redis<br />MongoDB<br />HTTPRequest<br />WebSocket<br />Amazon S3<br />And many others: <br /><br />
  30. gem install em-mysqlplus<br />EventMachine.rundo<br /> => 'localhost')<br /> query =conn.query("select sleep(1)")<br />query.callback { |res| pres.all_hashes }<br />query.errback { |res| pres.all_hashes }<br /> puts ”executing…”<br />end<br /># > ruby em-mysql-test.rb<br />#<br /># executing…<br /># [{"sleep(1)"=>"0"}]<br />callback fired 1s after “executing”<br />em-mysqlplus: example<br />asyncMySQL driver<br />
  31. non-blocking driver<br />require'mysqlplus'<br />defconnect(opts)<br />conn=connect_socket(opts)<br />, EventMachine::MySQLConnection, conn, opts, self)<br />end<br />defconnect_socket(opts)<br />conn=Mysql.init<br />conn.real_connect(host, user, pass, db, port, socket, ...)<br />conn.reconnect=false<br />conn<br />end<br /> reactor will poll & notify<br />em-mysqlplus: under the hood<br />mysqlplus + reactor loop<br />
  32. Features:<br /><ul><li> Maintains C-based mysql gem API
  33. Deferrables for every query with callback & errback
  34. Connection query queue - pile 'em up!
  35. Auto-reconnect on disconnects
  36. Auto-retry on deadlocks</li></ul><br />em-mysqlplus<br />mysqlplus + reactor loop<br />
  37. and this callback goes to…<br />
  38. We can do better than node.js<br />all the benefits of evented code without the drawbacks<br />
  39. Ruby 1.9 Fibers are a means of creating code blocks which can be paused and resumed by our application (think lightweight threads, minus the thread scheduler and less overhead). <br /> {<br />whiletruedo<br />Fiber.yield"Hi”<br />end<br />}<br />pf.resume# => Hi<br />pf.resume# => Hi<br />pf.resume# => Hi<br />Manual / cooperative scheduling!<br />Ruby 1.9 Fibers<br />and cooperative scheduling<br /><br />
  40. Fibers vs Threads: creation time much lower<br />Fibers vs Threads: memory usage is much lower<br />Ruby 1.9 Fibers<br />and cooperative scheduling<br /><br />
  41. defquery(sql)<br />f = Fiber.current<br /> => 'localhost')<br />q = conn.query(sql)<br />c.callback { f.resume(conn) }<br />c.errback { f.resume(conn) }<br />return Fiber.yield<br />end<br />EventMachine.rundo<br /> {<br /> res =query('select sleep(1)')<br /> puts "Results: #{res.fetch_row.first}"<br />}.resume<br />end<br />Exception, async!<br />Untangling Evented Code with Fibers<br /><br />
  42. defquery(sql)<br />f = Fiber.current<br /> => 'localhost')<br />q = conn.query(sql)<br />c.callback { f.resume(conn) }<br />c.errback { f.resume(conn) }<br /> return Fiber.yield<br />end<br />EventMachine.rundo<br />{<br /> res =query('select sleep(1)')<br /> puts "Results: #{res.fetch_row.first}"<br /> }.resume<br />end<br />1. Wrap into a continuation<br />Untangling Evented Code with Fibers<br /><br />
  43. defquery(sql)<br />f=Fiber.current<br /> => 'localhost')<br />q = conn.query(sql)<br />c.callback { f.resume(conn) }<br />c.errback { f.resume(conn) }<br />returnFiber.yield<br />end<br />EventMachine.rundo<br />{<br /> res =query('select sleep(1)')<br /> puts "Results: #{res.fetch_row.first}"<br /> }.resume<br />end<br />2. Pause the continuation<br />Untangling Evented Code with Fibers<br /><br />
  44. defquery(sql)<br />f=Fiber.current<br /> => 'localhost')<br />q = conn.query(sql)<br />c.callback { f.resume(conn) }<br />c.errback { f.resume(conn) }<br />returnFiber.yield<br />end<br />EventMachine.rundo<br />{<br /> res =query('select sleep(1)')<br /> puts "Results: #{res.fetch_row.first}"<br /> }.resume<br />end<br />3. Resume the continuation<br />Untangling Evented Code with Fibers<br /><br />
  45. Good news, you don’t even have to muck around with Fibers!<br />gem install em-synchrony<br /><br /><ul><li> Fiber aware connection pool with sync/async query support
  46. Multi request interface which accepts any callback enabled client
  47. Fibered iterator to allow concurrency control & mixing of sync / async
  48. em-http-request: .get, etc are synchronous, while .aget, etc are async
  49. em-mysqlplus: .query is synchronous, while .aquery is async
  50. remcached: .get, etc, and .multi_* methods are synchronous</li></ul>em-synchrony: simple evented programming<br />best of both worlds…<br />
  51. require"em-synchrony/em-mysqlplus"<br />EventMachine.synchronydo<br /> db"localhost")<br /> res =db.query("select sleep(1)")<br /> puts res<br />EventMachine.stop<br />end<br />Async under the hood<br />Untangling Evented Code with Fibers<br /><br />
  52. require "em-synchrony/em-mysqlplus"<br />EventMachine.synchrony do<br /> db = "localhost")<br /> res =db.query("select sleep(1)")<br /> puts res<br />EventMachine.stop<br />end<br />Untangling Evented Code with Fibers<br /><br />
  53. EM-HTTP, EM-MySQL, EM-Jack, etc.<br />Drivers<br />Async-rack<br />Ruby VM<br />Fibers<br />Network<br />Goliath<br />Thin<br />One VM, full concurrency, network-bound<br />Ruby 1.9, Fibers, Thin: in production!<br />
  54. git clone git://<br />git checkout activerecord<br />rake install<br />database.yml<br />development:<br />adapter:em_mysqlplus<br />database:widgets<br />pool: 5<br />timeout: 5000<br />environment.rb<br />require 'em-activerecord’<br />require 'rack/fiber_pool'<br /># Run each request in a Fiber<br />config.middleware.useRack::FiberPool<br />config.threadsafe!<br />Async Rails<br />with EventMachine & MySQL<br />
  55. classWidgetsController< ApplicationController<br />defindex<br />Widget.find_by_sql("select sleep(1)")<br />render:text => "Oh hai”<br />end<br />end<br />ab –c 5 –n 10<br />Server Software: thin<br />Server Hostname:<br />Server Port: 3000<br />Document Path: /widgets/<br />Document Length: 6 bytes<br />Concurrency Level: 5<br />Time taken for tests: 2.210 seconds<br />Complete requests: 10<br />Failed requests: 0<br />Requests per second: 4.53 [#/sec] (mean)<br />woot! Fiber DB pool at work.<br />Async Rails<br />with EventMachine & MySQL<br />
  56. git clone git://…./igrigorik/mysqlplus<br />git checkout activerecord<br />rake install<br />Not only is it doable… it already works.<br />
  57. Ruby 1.9 + Rails 3 + new stack<br />=<br />Order of magnitude better performance<br />(aka, enough of a reason to actually switch)<br />
  58. The state of art is not good enough, in fact, it’s terrible!<br />Let’s fix it.<br />Fibers & Cooperative Scheduling in Ruby:<br /><br />Untangling Evented Code with Ruby Fibers:<br /><br />EM-Synchrony:<br /><br />What do you think?<br />