CS378 Assignment 3

MPI programming (Due: April 8^th, 2008)

This assignment requires you to install and run MPI benchmarks to evaluate the latencies of MPI operations. Get started early, since installation and running benchmarks might be rather involved.

Problem 1: Run the IMB benchmark package described below to determine the latencies of various MPI operations:

- Point to Point: (with 2 processors)

o PingPong

o PingPing

- Parallel Transfer (with 4 processors)

o SendRecv

o Exchange

- Collective (with 4 processors)

o Bcast

o Allgather

o Alltoall

o Reduce

o AllReduce

o Barrier

Turn  in  the output from your program.

Problem 2:  Modern distributed-memory machines are built from multicore processors, so the communication latency between two processes on the same

multicore package can be substantially less than the communication latency between processes on different packages. In this problem, you are required to measure the
communication latencies for different message sizes for both scenarios as follows.

• Your program should create 16 processes.
  
• It should prompt the user for the ranks of the sending and receiving processes (call these s and r). If s and r are zero, the program terminates.
  
• Otherwise, the program uses blocking sends and receives to exchange messages between s and r some number of times, and measures the average communication latency.
• It should do this for a range of message sizes (say from 1 double to 10,000 doubles) .
• It then prompts the user again for the ranks of the sending and receiving processes etc.
  

Using the data from your experiment, answer the following questions.

• Plot a graph of the latency of point-to-point communication vs. message size if s and r are 0 and 1 respectively. What is the start-up cost and  cost/byte of the communication in this case?
• Repeat this for s = 0 and r = 31.

Submit your code, the graphs and the answers to the questions.

Problem 3: In class, we discussed a program that uses numerical integration to estimate the value of pi. Modify this program to accomplish the following.

• Your program should create 32 processes.
• It should prompt the user for the number of processes that should be used to perform the computation. If this number is zero, the program terminates.
• Otherwise, it performs the numerical integration on that number of processes, and measures the time required to perform that computation.

Using this data, plot a graph of speed-up vs. the number of processes used to perform the numerical integration. Submit your code and this graph.

*******************************************************************************************************************************

Docs/Files/Login required for the assignment:

• IMB benchmarks containing

1. Readme_first

2. doc/Readme_IMB.txt

3. doc/IMB_ug-3.0.pdf

4. All sources for the benchmarks.

• lonestar-batch-system.ppt
• A login to lonestar.tacc.utexas.edu (You should receive a mail with login and password.)

Steps for installing and running the benchmarks:

1) scp IMB_3.0.tar.gz username@lonestar.tacc.utexas.edu:

2) tar –zxvf IMB_3.0.tar.gz

3) cd IMB_3.0

4) Follow instructions in Readme_First, which directs the user to read:

a.  ./doc/Readme_IMB.txt which contains the installation information.

b. ./doc/IMB_ug.pdf which contains the detailed information about each benchmark.

The condensed installation instructions are given in the next step.

5) Installation:

a. Set the path of the variable MPI_HOME in ./src/makefile_mpich, the path being obtained by running “which mpicc” :

i. MPI_HOME=/opt/MPI/intel9/mvapich-gen2/0.9.8/

b. cd src

c. make –f make_mpich

6) Running:

Step 5 generates an executable IMB-MPI1 in the src directory, which can be run using the instructions mentioned in IMB_ug.pdf Section 5.1.2

a. Before running, a textfile containing a list of machines on which the tests can be run. This needs to be passed in the “-machinefile” option of IMB-MPI1. A sample machinefile is given at the end of the doc. It has “lonestar.tacc.utexas.edu” repeated 5 times with newlines separating each entry.

b. To run the Barrier benchmark for instance, Create a jobscript (at the end of the doc. See lonestar-batch-system.ppt for a description of how to create jobscripts)

c. Run the jobscript in the directory where IMB-MPI1 is present by using “bsub < jobscript” (See lonestar-batch-system.ppt for an explanation of what bsub does)

d. Change parameters to ibrun to run different benchmarks. See IMB_ug.pdf for an explanation of various benchmarks, relevant parameters and such.

Sample jobscript for “Barrier Benchmark”:

#!/bin/csh

#BSUB -n 4

#BSUB -J Barrier

#BSUB -o %J.out

#BSUB -e %J.err

#BSUB -q normal

#BSUB -W 0:15

#BSUB -u vishwasm@cs.utexas.edu

#BSUB -B

#BSUB -N

echo "Master Host = "`hostname`

echo "LSF_SUBMIT_DIR: $LS_SUBCWD"

echo "PWD_DIR: "`pwd`

ibrun  IMB-MPI1 -npmin 4 -machinefile hostfile Barrier

Sample machinefile:

lonestar.tacc.utexas.edu

lonestar.tacc.utexas.edu

lonestar.tacc.utexas.edu

lonestar.tacc.utexas.edu

lonestar.tacc.utexas.edu

Some important Files required for running:

bsub

ibrun

mpicc

IMB-MPI1 (created by make_mpich)

machinefile (to be created)

jobscript (to be created)