1. 1b.1 Types of Parallel Computers Two principal approaches: Shared memory multiprocessor Distributed memory multicomputer ITCS 4/5145 Parallel Programming, UNC-Charlotte, B. Wilkinson, 2012. Jan 14, 2013

2. 1b.2 Shared Memory Multiprocessor

3. 1b.3 Conventional Computer Consists of a processor executing a program stored in a (main) memory: Each main memory location located by its address. Addresses start at 0 and extend to 2 b - 1 when there are b bits (binary digits) in address. Main memory Processor Instructions (to processor) Data (to or from processor)

4. 1b.4 Shared Memory Multiprocessor System Natural way to extend single processor model - have multiple processors connected to multiple memory modules, such that each processor can access any memory module: Processors Processor-memory Interconnections Memory module One address space

5. 1b.5 Simplistic view of a small shared memory multiprocessor Examples: Dual Pentiums Quad Pentiums ProcessorsShared memory Bus

6. 1b.6 Real computer system have cache memory between main memory and processors. Level 1 (L1) cache and Level 2 (L2) cache. Example Quad Shared Memory Multiprocessor Processor L2 Cache Bus interface L1 cache Processor L2 Cache Bus interface L1 cache Processor L2 Cache Bus interface L1 cache Processor L2 Cache Bus interface L1 cache Memory controller Memory Processor/ memory bus Shared memory

7. 1b.7 “Recent” innovation (since 2005) Dual-core and multi-core processors Two or more independent processors in one package Actually an old idea but not put into wide practice until recently with the limits of making single processors faster principally caused by: –Power dissipation (power wall) and clock frequency limitations –Limits in parallelism within a single instruction stream –Memory speed limitations (memory wall)

8. 1b.8 “The Free Lunch Is Over: A Fundamental Turn Toward Concurrency in Software” Herb Sutter, http://www.gotw.ca/publications/concurrency-ddj.htm Power dissipation Clock frequency

9. 1b.9 Single “quad core” shared memory multiprocessor L2 Cache Memory controller Memory Shared memory Chip Processor L1 cache Processor L1 cache Processor L1 cache Processor L1 cache

10. 1b.10 Multiple quad-core multiprocessors (example coit-grid05.uncc.edu) Memory controller Memory Shared memory L2 Cache possible L3 cache Processor L1 cache Processor L1 cache Processor L1 cache Processor L1 cache Processor L1 cache Processor L1 cache Processor L1 cache Processor L1 cache

11. 1b.11 Programming Shared Memory Multiprocessors Several possible ways – we will concentrate upon using threads Threads - individual parallel sequences (threads), each thread having their own local variables but being able to access shared variables declared outside threads. 1.Low–level thread libraries - programmer calls thread routines to create and control the threads. Example Pthreads, Java threads. 2.Higher level library functions and preprocessor compiler directives. Example OpenMP - industry standard. Consists of library functions, compiler directives, and environment variables

12. 1b.12 Tasks – rather than program with threads, which are closely linked to the physical hardware, can program with parallel “tasks”. Promoted by Intel with their TBB (Thread Building Blocks) tools. Other alternatives include parallelizing compilers compiling regular sequential programs and making them parallel programs, and special parallel languages (both not now common).

13. GPU clusters Recent trend for clusters – incorporating GPUs for high performance. GPU often attached through PCI-e x16 interface to CPU, and separate GPU memory. Now 1000’s cores in each GPU offering orders of magnitude speed improvement for HPC tasks. 10,000’s of threads possible (Data parallel programming model, see later) 1b.13

14. 1b.14 Message-Passing Multicomputer

15. 1b.15 Message-Passing Multicomputer Complete computers connected through an interconnection network: Processor Interconnection network Local Computers Messages memory Many interconnection networks explored in the 1970s and 1980s including 2- and 3- dimensional meshes, hypercubes, and multistage interconnection networks

16. 1b.16 Networked Computers as a Computing Platform Became a very attractive alternative to expensive supercomputers and parallel computer systems for high-performance computing in early 1990s. Several early projects. Notable: – Berkeley NOW (network of workstations) project. –NASA Beowulf project.

17. 1b.17 Key advantages: Very high performance workstations and PCs readily available at low cost. The latest processors can easily be incorporated into the system as they become available. Existing software can be used or modified.

18. 1b.18 Beowulf Clusters A group of interconnected “commodity” computers achieving high performance with low cost. Typically using commodity interconnects - high speed Ethernet, and Linux OS. Beowulf comes from name given by NASA Goddard Space Flight Center cluster project.

19. 1b.19 Cluster Interconnects Originally fast Ethernet on low cost clusters Gigabit Ethernet - easy upgrade path More specialized/higher performance interconnects available including Myrinet and Infiniband.

20. 1b.20 Dedicated cluster with a master node and compute nodes User Master node Compute nodes Dedicated Cluster Ethernet interface Switch External network Computers Local network

21. 1b.21 Software Tools for Clusters Based upon message passing programming model User-level libraries provided for explicitly specifying messages to be sent between executing processes on each computer. Use with regular programming languages (C, C++,...). Can be quite difficult to program correctly as we shall see.

22. Next step Learn the message passing programming model, some MPI routines, write a message-passing program and test on the cluster. 1b.22