1. Mass Data Processing Technology on Large Scale Clusters Summer, 2007, Tsinghua University All course material (slides, labs, etc) is licensed under the Creative Commons Attribution 2.5 License.Creative Commons Attribution 2.5 License Many thanks to Aaron Kimball & Sierra Michels-Slettvet for their original version 1

2. Introduction to Distributed SystemsParallelization & SynchronizationFundamentals of NetworkingRemote Procedure Calls (RPC)Transaction Processing Systems 2

3. 3

4. Computer Speedup 4 Moore’s Law: “The density of transistors on a chip doubles every 18 months, for the same cost” (1965) Image: Tom’s Hardware Why slow down here? Then, How to improve the performance?

5. Scope of Problems What can you do with 1 computer? What can you do with 100 computers? What can you do with an entire data center? 5

6. Distributed Problems Rendering multiple frames of high-quality animation 6 Image: DreamWorks Animation

7. Distributed Problems Simulating several hundred or thousand characters 7 Happy Feet © Kingdom Feature Productions; Lord of the Rings © New Line Cinema

8. Distributed Problems Indexing the web (Google) Simulating an Internet-sized network for networking experiments (PlanetLab) Speeding up content delivery (Akamai) 8 What is the key attribute that all these examples have in common?

9. PlanetLab 9 PlanetLab is a global research network that supports the development of new network services. PlanetLab currently consists of 809 nodes at 401 sites.

10. CDN - Akamai 10

11. Parallel vs. Distributed Parallel computing can mean: Vector processing of data (SIMD) Multiple CPUs in a single computer (MIMD) Distributed computing is multiple CPUs across many computers (MIMD) 11

12. A Brief History… 1975-85 Parallel computing was favored in the early years Primarily vector-based at first Gradually more thread- based parallelism was introduced 12 Cray 2 supercomputer (Wikipedia)

13. A Brief History… 1985-95 “Massively parallel architectures” start rising in prominence Message Passing Interface (MPI) and other libraries developed Bandwidth was a big problem 13

14. A Brief History… 1995-Today Cluster/grid architecture increasingly dominant Special node machines eschewed in favor of COTS technologies Web-wide cluster software Companies like Google take this to the extreme (10,000 node clusters) 14

15. Top 500, Architecture 15

16. 16

17. Parallelization Idea Parallelization is “easy” if processing can be cleanly split into n units: 17

18. Parallelization Idea (2) 18 In a parallel computation, we would like to have as many threads as we have processors. e.g., a four- processor computer would be able to run four threads at the same time.

19. Parallelization Idea (3) 19

20. Parallelization Idea (4) 20

21. Parallelization Pitfalls But this model is too simple! How do we assign work units to worker threads? What if we have more work units than threads? How do we aggregate the results at the end? How do we know all the workers have finished? What if the work cannot be divided into completely separate tasks? 21 What is the common theme of all of these problems?

22. Parallelization Pitfalls (2) Each of these problems represents a point at which multiple threads must communicate with one another, or access a shared resource. Golden rule: Any memory that can be used by multiple threads must have an associated synchronization system! 22

23. What’s Wrong With This? 23 Thread 1: void foo() { x++; y = x; } Thread 2: void bar() { y++; x++; } If the initial state is y = 0, x = 6, what happens after these threads finish running?

24. Multithreaded = Unpredictability 24 When we run a multithreaded program, we don’t know what order threads run in, nor do we know when they will interrupt one another. Thread 1: void foo() { eax = mem[x]; inc eax; mem[x] = eax; ebx = mem[x]; mem[y] = ebx; } Thread 2: void bar() { eax = mem[y]; inc eax; mem[y] = eax; eax = mem[x]; inc eax; mem[x] = eax; } Many things that look like “one step” operations actually take several steps under the hood:

25. Multithreaded = Unpredictability This applies to more than just integers: Pulling work units from a queue Reporting work back to master unit Telling another thread that it can begin the “next phase” of processing … All require synchronization! 25

26. Synchronization Primitives A synchronization primitive is a special shared variable that guarantees that it can only be accessed atomically. Hardware support guarantees that operations on synchronization primitives only ever take one step 26

27. Semaphores A semaphore is a flag that can be raised or lowered in one step Semaphores were flags that railroad engineers would use when entering a shared track 27 Only one side of the semaphore can ever be red! (Can both be green?)

28. Semaphores set() and reset() can be thought of as lock() and unlock() Calls to lock() when the semaphore is already locked cause the thread to block. Pitfalls: Must “bind” semaphores to particular objects; must remember to unlock correctly 28

29. The “corrected” example 29 Thread 1: void foo() { sem.lock(); x++; y = x; sem.unlock(); } Thread 2: void bar() { sem.lock(); y++; x++; sem.unlock(); } Global var “Semaphore sem = new Semaphore();” guards access to x & y

30. Condition Variables A condition variable notifies threads that a particular condition has been met Inform another thread that a queue now contains elements to pull from (or that it’s empty – request more elements!) Pitfall: What if nobody’s listening? 30

31. The Final Example 31 Thread 1: void foo() { sem.lock(); x++; y = x; fooDone = true; sem.unlock(); fooFinishedCV.notify(); } Thread 2: void bar() { sem.lock(); if(!fooDone) fooFinishedCV.wait(sem); y++; x++; sem.unlock(); } Global vars: Semaphore sem = new Semaphore(); ConditionVar fooFinishedCV = new ConditionVar(); boolean fooDone = false;

32. Barriers A barrier knows in advance how many threads it should wait for. Threads “register” with the barrier when they reach it, and fall asleep. Barrier wakes up all registered threads when total count is correct Pitfall: What happens if a thread takes a long time? 32 Barrier

33. Too Much Synchronization? Deadlock 33 Synchronization becomes even more complicated when multiple locks can be used Can cause entire system to “get stuck” Thread A: semaphore1.lock(); semaphore2.lock(); /* use data guarded by semaphores */ semaphore1.unlock(); semaphore2.unlock(); Thread B: semaphore2.lock(); semaphore1.lock(); /* use data guarded by semaphores */ semaphore1.unlock(); semaphore2.unlock(); (Image: RPI CSCI.4210 Operating Systems notes)

34. The Moral: Be Careful! Synchronization is hard Need to consider all possible shared state Must keep locks organized and use them consistently and correctly Knowing there are bugs may be tricky; fixing them can be even worse! Keeping shared state to a minimum reduces total system complexity 34

35. 35

36. Sockets: The Internet = tubes? A socket is the basic network interface Provides a two-way “pipe” abstraction between two applications Client creates a socket, and connects to the server, who receives a socket representing the other side 36

37. Ports Within an IP address, a port is a sub-address identifying a listening program Allows multiple clients to connect to a server at once 37

38. Example: Web Server (1/3) 38 The server creates a listener socket attached to a specific port. 80 is the agreed-upon port number for web traffic.

39. Example: Web Server (2/3) 39 The client-side socket is still connected to a port, but the OS chooses a random unused port number When the client requests a URL (e.g., “www.google.com”), its OS uses a system called DNS to find its IP address.

40. Example: Web Server (3/3) 40 Listener is ready for more incoming connections, while we process the current connection in parallel

41. Example: Web Server 41

42. What makes this work? Underneath the socket layer are several more protocols Most important are TCP and IP (which are used hand-in- hand so often, they’re often spoken of as one protocol: TCP/IP) 42 Even more low-level protocols handle how data is sent over Ethernet wires, or how bits are sent through the air using 802.11 wireless…

43. IP: The Internet Protocol Defines the addressing scheme for computers Encapsulates internal data in a “packet” Does not provide reliability Just includes enough information for the data to tell routers where to send it 43

44. TCP: Transmission Control Protocol Built on top of IP Introduces concept of “connection” Provides reliability and ordering 44

45. Why is This Necessary? Not actually tube-like “underneath the hood” Unlike phone system (circuit switched), the packet switched Internet uses many routes at once 45

46. Networking Issues If a party to a socket disconnects, how much data did they receive? … Did they crash? Or did a machine in the middle? Can someone in the middle intercept/modify our data? Traffic congestion makes switch/router topology important for efficient throughput 46

47. 47

48. How RPC Doesn’t Work Regular client-server protocols involve sending data back and forth according to a shared state 48 Client: Server: HTTP/1.0 index.html GET 200 OK Length: 2400 (file data) HTTP/1.0 hello.gif GET 200 OK Length: 81494 …

49. Remote Procedure Call RPC servers will call arbitrary functions in dll, exe, with arguments passed over the network, and return values back over network 49 Client: Server: foo.dll,bar(4, 10, “hello”) “returned_string” foo.dll,baz(42) err: no such function …

50. Possible Interfaces RPC can be used with two basic interfaces: synchronous and asynchronous Synchronous RPC is a “remote function call” – client blocks and waits for return val Asynchronous RPC is a “remote thread spawn” 50

51. Synchronous RPC 51

52. Asynchronous RPC 52

53. Asynchronous RPC 2: Callbacks 53

54. Wrapper Functions Writing rpc_call(foo.dll, bar, arg0, arg1..) is poor form Confusing code Breaks abstraction Wrapper function makes code cleaner bar(arg0, arg1); //just write this; calls “stub” 54

55. More Design Considerations Who can call RPC functions? Anybody? How do you handle multiple versions of a function? Need to marshal objects How do you handle error conditions? Numerous protocols: DCOM, CORBA, JRMI… 55

56. 56

57. TPS: Definition A system that handles transactions coming from several sources concurrently Transactions are “events that generate and modify data stored in an information system for later retrieval” * 57 * http://en.wikipedia.org/wiki/Transaction_Processing_System

58. Key Features of TPS: ACID “ACID” is the acronym for the features a TPS must support: Atomicity – A set of changes must all succeed or all fail Consistency – Changes to data must leave the data in a valid state when the full change set is applied Isolation – The effects of a transaction must not be visible until the entire transaction is complete Durability – After a transaction has been committed successfully, the state change must be permanent. 58

59. Atomicity & Durability What happens if we write half of a transaction to disk and the power goes out? 59

60. Logging: The Undo Buffer 1. Database writes to log the current values of all cells it is going to overwrite 2. Database overwrites cells with new values 3. Database marks log entry as committed If db crashes during (2), we use the log to roll back the tables to prior state 60

61. Consistency: Data Types Data entered in databases have rigorous data types associated with them, and explicit ranges Does not protect against all errors (entering a date in the past is still a valid date, etc), but eliminates tedious programmer concerns 61

62. Consistency: Foreign Keys Database designers declare that fields are indices into the keys of another table Database ensures that target key exists before allowing value in source field 62

63. Isolation Using mutual-exclusion locks, we can prevent other processes from reading data we are in the process of writing When a database is prepared to commit a set of changes, it locks any records it is going to update before making the changes 63

64. Faulty Locking Locking alone does not ensure isolation! Changes to table A are visible before changes to table B – this is not an isolated transaction 64

65. Two-Phase Locking After a transaction has released any locks, it may not acquire any new locks Effect: The lock set owned by a transaction has a “growing” phase and a “shrinking” phase 65

66. Relationship to Distributed Comp At the heart of a TPS is usually a large database server Several distributed clients may connect to this server at points in time Database may be spread across multiple servers, but must still maintain ACID 66

67. Conclusions We’ve seen 3 layers that make up a distributed system Designing a large distributed system involves engineering tradeoffs at each of these levels Appreciating subtle concerns at each level requires diving past the abstractions, but abstractions are still useful in general 67

68. Discussion Distributed System Design 68

69. Questions and Answers !!! 69