1. MPI and C-Language Seminars 2010

2. Seminar Plan  Week 1 – Introduction, Data Types, Control Flow, Pointers  Week 2 – Arrays, Structures, Enums, I/O, Memory  Week 3 – Compiler Options and Debugging  Week 4 – MPI in C and Using the HPSG Cluster  Week 5 – “How to Build a Performance Model”  Week 6-9 – Coursework Troubleshooting (Seminar tutors available in their office)

4. Performance Models  Aim to predict the runtime of an application.  Allow you to predict beyond grid size and unknown hardware.  Gauge the scalability of the algorithm / code.  Help analyse the parallel efficiency of code.  See where bottle necks are. Both hardware and software.

5. Factors of a Model  Computation – Active Processing: The time spent doing actual work. More processors => less work per processor. Overall computation time should fall with increase in processors.  Communication – Message Passing: Communication between processors. Overall I/O time will increase with processor count. More network contention means slower communication.

6. Getting the Balance (1 / 2) Number Of Processors

7. Getting the Balance (2 / 2)  Fixed Costs: The work done by all processors. More processors will not reduce this time.  Variable Costs: Portion of work which varies with processor count. Generally based on problem size decomposition.

8. Timers  Lots of different timers: CPU Time – Actual time spent of CPU. Wall Time – Total time since program start.  Different timers have different overheads.  Try and avoid timing timers.  Recommend C timer – Need to call with 2 double pointers. double cpuStart, wallStart; Timers(&wallStart, &cpuStart);

10. Where is the Expense?  Need to establish what the expensive operations are: Functions which are called frequently. Functions which take a long time. Work out a percentage break down of total time.  Is it communicational or computational expense?  Is it a fixed cost or a variable cost?

11. Computational Model (1 / 2)  How will the number of processors effect the amount of work done by each processor.  Will they all the same amount of work? Even decomposition.  Are loops dependent on the problem size?  Need to look at: How long operations take. How many times they are performed.

12. Computational Model (2 / 2)  A basic model: Time how long each different operation takes. Calculate how many times those operations are used. Add them all up.  Inaccuracy: When using timers consider their overhead. Always more accurate to time the repetition of operations and divide through.  Note: Communication will show in wall time.

13. Communication Model (1 / 2)  Many different types of communication: Send and receives. Collective operations. Barriers.  Need to build a model of the network: Can use existing programs: PingPong / SKaMPI Or write your own.  How much data is being sent?

14. Communication Model (2 / 2)  Communication times are based on packet size. There is an initial cost of a send – Handshake. Then a variable cost – Payload.  Where is the data being sent? Are the source and destination on the same node? Message Size (Bytes)

15. Bringing it all Together  What you need: Computation benchmark application. Communication benchmark application. Spreadsheet model.  Run benchmarks on cluster and plug data into model.  Make predictions for different processor configurations.