1. Networks of Workstations Prabhaker Mateti Wright State University

2. Prabhaker Mateti, Networks of Workstations 2 Overview Parallel computers Concurrent computation Parallel Methods Message Passing Distributed Shared Memory Programming Tools Cluster configurations

3. Prabhaker Mateti, Networks of Workstations 3 Granularity of Parallelism Fine-Grained Parallelism Medium-Grained Parallelism Coarse-Grained Parallelism NOWs (Networks of Workstations)

4. Prabhaker Mateti, Networks of Workstations 4 Fine-Grained Machines Tens of thousands of Processors Processors –Slow (bit serial) –Small (K bits of RAM) –Distributed Memory Interconnection Networks –Message Passing Single Instruction Multiple Data (SIMD)

5. Prabhaker Mateti, Networks of Workstations 5 Sample Meshes Massively Parallel Processor (MPP) TMC CM-2 (Connection Machine) MasPar MP-1/2

6. Prabhaker Mateti, Networks of Workstations 6 Medium-Grained Machines Typical Configurations –Thousands of processors –Processors have power between coarse- and fine-grained Either shared or distributed memory Traditionally: Research Machines Single Code Multiple Data (SCMD)

7. Prabhaker Mateti, Networks of Workstations 7 Medium-Grained Machines Ex: Cray T3E Processors –DEC Alpha EV5 (600 MFLOPS peak) –Max of 2048 Peak Performance: 1.2 TFLOPS 3-D Torus Memory: 64 MB - 2 GB per CPU

8. Prabhaker Mateti, Networks of Workstations 8 Coarse-Grained Machines Typical Configurations –Hundreds of Processors Processors –Powerful (fast CPUs) –Large (cache, vectors, multiple fast buses) Memory: Shared or Distributed-Shared Multiple Instruction Multiple Data (MIMD)

9. Prabhaker Mateti, Networks of Workstations 9 Coarse-Grained Machines SGI Origin 2000: –PEs (MIPS R10000): Max of 128 –Peak Performance: 49 Gflops –Memory: 256 GBytes –Crossbar switches for interconnect HP/Convex Exemplar: –PEs (HP PA-RISC 8000): Max of 64 –Peak Performance: 46 Gflops –Memory: Max of 64 GBytes –Distributed crossbar switches for interconnect

10. Prabhaker Mateti, Networks of Workstations 10 Networks of Workstations Exploit inexpensive Workstations/PCs Commodity network The NOW becomes a “distributed memory multiprocessor” Workstations send+receive messages C and Fortran programs with PVM, MPI, etc. libraries Programs developed on NOWs are portable to supercomputers for production runs

11. Prabhaker Mateti, Networks of Workstations 11 “Parallel” Computing Concurrent Computing Distributed Computing Networked Computing Parallel Computing

12. Prabhaker Mateti, Networks of Workstations 12 Definition of “Parallel” S1 begins at time b1, ends at e1 S2 begins at time b2, ends at e2 S1 || S2 –Begins at min(b1, b2) –Ends at max(e1, e2) –Equiv to S2 || S1

13. Prabhaker Mateti, Networks of Workstations 13 Data Dependency x := a + b; y := c + d; x := a + b || y := c + d; y := c + d; x := a + b; X depends on a and b, y depends on c and d Assumed a, b, c, d were independent

14. Prabhaker Mateti, Networks of Workstations 14 Types of Parallelism Result Specialist Agenda

15. Prabhaker Mateti, Networks of Workstations 15 Perfect Parallelism Also called –Embarrassingly Parallel –Result parallel Computations that can be subdivided into sets of independent tasks that require little or no communication –Monte Carlo simulations –F(x, y, z)

16. Prabhaker Mateti, Networks of Workstations 16 MW Model Manager –Initiates computation –Tracks progress –Handles worker’s requests –Interfaces with user Workers –Spawned and terminated by manager –Make requests to manager –Send results to manager

17. Prabhaker Mateti, Networks of Workstations 17 Reduction Combine several sub-results into one Reduce r1 r2 … rn with op Becomes r1 op r2 op … op rn

18. Prabhaker Mateti, Networks of Workstations 18 Data Parallelism Also called –Domain Decomposition –Specialist Same operations performed on many data elements simultaneously –Matrix operations –Compiling several files

19. Prabhaker Mateti, Networks of Workstations 19 Control Parallelism Different operations performed simultaneously on different processors E.g., Simulating a chemical plant; one processor simulates the preprocessing of chemicals, one simulates reactions in first batch, another simulates refining the products, etc.

20. Prabhaker Mateti, Networks of Workstations 20 Process communication Shared Memory Message Passing

21. Prabhaker Mateti, Networks of Workstations 21 Shared Memory Process A writes to a memory location Process B reads from that memory location Synchronization is crucial Excellent speed

22. Prabhaker Mateti, Networks of Workstations 22 Shared Memory Needs hardware support: –multi-ported memory Atomic operations: –Test-and-Set –Semaphores

23. Prabhaker Mateti, Networks of Workstations 23 Shared Memory Semantics: Assumptions Global time is available. Discrete increments. Shared variable s, = vi at ti, i=0,… Process A: s := v1 at time t1 Assume no other assignment occurred after t1. Process B reads s at time t and gets value v.

24. Prabhaker Mateti, Networks of Workstations 24 Shared Memory: Semantics Value of Shared Variable –v = v1, if t > t1 –v = v0, if t < t1 –v = ??, if t = t1 t = t1 +- discrete quantum Next Update of Shared Variable –Occurs at t2 –t2 = t1 + ?

25. Prabhaker Mateti, Networks of Workstations 25 Condition Variables and Semaphores Semaphores –V(s) ::= –P(s) ::= 0 do s := s – 1> Condition variables –C.wait() –C.signal()

26. Prabhaker Mateti, Networks of Workstations 26 Distributed Shared Memory A common address space that all the computers in the cluster share. Difficult to describe semantics.

27. Prabhaker Mateti, Networks of Workstations 27 Distributed Shared Memory: Issues Distributed –Spatially –LAN –WAN No global time available

28. Prabhaker Mateti, Networks of Workstations 28 Messages Messages are sequences of bytes moving between processes The sender and receiver must agree on the type structure of values in the message “Marshalling” of data

29. Prabhaker Mateti, Networks of Workstations 29 Message Passing Process A sends a data buffer as a message to process B. Process B waits for a message from A, and when it arrives copies it into its own local memory. No memory shared between A and B.

30. Prabhaker Mateti, Networks of Workstations 30 Message Passing Obviously, –Messages cannot be received before they are sent. –A receiver waits until there is a message. Asynchronous –Sender never blocks, even if infinitely many messages are waiting to be received –Semi-asynchronous is a practical version of above with large but finite amount of buffering

31. Prabhaker Mateti, Networks of Workstations 31 Message Passing: Point to Point Q: send(m, P) –Send message M to process P P: recv(x, Q) –Receive message from process P, and place it in variable x The message data –Type of x must match that of m – As if x := m

32. Prabhaker Mateti, Networks of Workstations 32 Broadcast One sender, multiple receivers Not all receivers may receive at the same time

33. Prabhaker Mateti, Networks of Workstations 33 Types of Sends Synchronous Asynchronous

34. Prabhaker Mateti, Networks of Workstations 34 Synchronous Message Passing Sender blocks until receiver is ready to receive. Cannot send messages to self. No buffering.

35. Prabhaker Mateti, Networks of Workstations 35 Message Passing: Speed Speed not so good –Sender copies message into system buffers. –Message travels the network. –Receiver copies message from system buffers into local memory. –Special virtual memory techniques help.

36. Prabhaker Mateti, Networks of Workstations 36 Message Passing: Programming Less error-prone cf. shared memory

37. Prabhaker Mateti, Networks of Workstations 37 Message Passing: Synchronization Synchronous MP: –Sender waits until receiver is ready. –No intermediary buffering

38. Prabhaker Mateti, Networks of Workstations 38 Barrier Synchronization Processes wait until “all” arrive

39. Prabhaker Mateti, Networks of Workstations 39 Parallel Software Development Algorithmic conversion by compilers

40. Prabhaker Mateti, Networks of Workstations 40 Development of Distributed+Parallel Programs New code + algorithms Old programs rewritten –in new languages that have distributed and parallel primitives –With new libraries Parallelize legacy code

41. Prabhaker Mateti, Networks of Workstations 41 Conversion of Legacy Software Mechanical conversion by software tools Reverse engineer its design, and re-code

42. Prabhaker Mateti, Networks of Workstations 42 Automatically parallelizing compilers Compilers analyze programs and parallelize (usually loops). Easy to use, but with limited success

43. Prabhaker Mateti, Networks of Workstations 43 OpenMP on Networks of Workstations The OpenMP is an API for shared memory architectures. User-gives hints as directives to the compiler http://www.openmp.org

44. Prabhaker Mateti, Networks of Workstations 44 Message Passing Libraries Programmer is responsible for data distribution, synchronizations, and sending and receiving information Parallel Virtual Machine (PVM) Message Passing Interface (MPI) BSP

45. Prabhaker Mateti, Networks of Workstations 45 BSP: Bulk Synchronous Parallel model Divides computation into supersteps In each superstep a processor can work on local data and send messages. At the end of the superstep, a barrier synchronization takes place and all processors receive the messages which were sent in the previous superstep http://www.bsp-worldwide.org/

46. Prabhaker Mateti, Networks of Workstations 46 BSP Library Small number of subroutines to implement –process creation, –remote data access, and –bulk synchronization. Linked to C, Fortran, … programs

47. Prabhaker Mateti, Networks of Workstations 47 Parallel Languages Shared-memory languages Parallel object-oriented languages Parallel functional languages Concurrent logic languages

48. Prabhaker Mateti, Networks of Workstations 48 Tuple Space: Linda Atomic Primitives –In (t) –Read (t) –Out (t) –Eval (t) Host language: e.g., JavaSpaces

49. Prabhaker Mateti, Networks of Workstations 49 Data Parallel Languages Data is distributed over the processors as a arrays Entire arrays are manipulated: –A(1:100) = B(1:100) + C(1:100) Compiler generates parallel code –Fortran 90 –High Performance Fortran (HPF)

50. Prabhaker Mateti, Networks of Workstations 50 Parallel Functional Languages Erlang http://www.erlang.org/ http://www.erlang.org/ SISAL http://www.llnl.gov/sisal/http://www.llnl.gov/sisal/ PCN Argonne

51. Prabhaker Mateti, Networks of Workstations 51 Clusters

52. Prabhaker Mateti, Networks of Workstations 52 Buildings-Full of Workstations 1.Distributed OS have not taken a foot hold. 2.Powerful personal computers are ubiquitous. 3.Mostly idle: more than 90% of the up- time? 4.100 Mb/s LANs are common. 5.Windows and Linux are the top two OS in terms of installed base.

53. Prabhaker Mateti, Networks of Workstations 53 Cluster Configurations NOW -- Networks of Workstations COW -- Clusters of Dedicated Nodes Clusters of Come-and-Go Nodes Beowulf clusters

54. Prabhaker Mateti, Networks of Workstations 54 Beowulf Collection of compute nodes Full trust in each other –Login from one node into another without authentication –Shared file system subtree Dedicated

55. Prabhaker Mateti, Networks of Workstations 55 Close Cluster Configuration compute node compute node compute node compute node Front-end High Speed Network Service Network gateway node External Network compute node compute node compute node compute node Front-end High Speed Network gateway node External Network File Server node

56. Prabhaker Mateti, Networks of Workstations 56 Open Cluster Configuration compute node compute node compute node compute node compute node compute node compute node compute node Front-end External Network File Server node High Speed Network

57. Prabhaker Mateti, Networks of Workstations 57 Interconnection Network Most popular: Fast Ethernet Network topologies –Mesh –Torus Switch v Hub

58. Prabhaker Mateti, Networks of Workstations 58 Software Components Operating System –Linux, FreeBSD, … Parallel programming –PVM, MPI Utilities, … Open source

59. Prabhaker Mateti, Networks of Workstations 59 Software Structure of PC Cluster HIGH-SPEED NETWORK HARDWARE LAYER OS LAYER HARDWARE LAYER OS LAYER HARDWARE LAYER OS LAYER PARALLEL VIRTUAL MACHINE LAYER Parallel Program

60. Prabhaker Mateti, Networks of Workstations 60 Single System View Single system view –Common filesystem structure view from any node –Common accounts on all nodes –Single software installation point Benefits –Easy to install and maintain system –Easy to use for users

61. Prabhaker Mateti, Networks of Workstations 61 Installation Steps Install Operating system Setup a Single System View –Shared filesystem –Common accounts –Single software installation point Install parallel programming packages such as MPI, PVM, BSP Install utilities, libraries, and applications

62. Prabhaker Mateti, Networks of Workstations 62 Linux Installation Linux has many distributions: Redhat, Caldera, SuSe, Debian, … Caldera is easy to install All above upgrade with RPM package management Mandrake and SuSe come with a very complete set of software

63. Prabhaker Mateti, Networks of Workstations 63 Clusters with Part Time Nodes Cycle Stealing: Running of jobs on a workstation that don't belong to the owner. Definition of Idleness: E.g., No keyboard and no mouse activity Tools/Libraries –Condor –PVM –MPI

64. Prabhaker Mateti, Networks of Workstations 64 Migration of Jobs Policies –Immediate-Eviction –Pause-and-Migrate Technical Issues –Checkpointing: Preserving the state of the process so it can be resumed. –Migrating from one architecture to another

65. Prabhaker Mateti, Networks of Workstations 65 Summary Parallel –computers –computation Parallel Methods Communication primitives –Message Passing –Distributed Shared Memory Programming Tools Cluster configurations