Revisions to PA14 for the final report

• Verify the MPI process scheduling policy on the OS cluster and HPC cluster. Here is a program hello.c that prints the hostname of the node where each MPI process resides, and verify the difference between different scheduling policies by doing a pair-wise ping-pong message test. (The instructor wrote the program and tested it on the OS cluster. It seems that by repeating each node name in the machinefile the neighboring MPI processes are indeed assigned to the same node. In this case, the ping-pong throughput is considerably higher than scheduling them to different nodes. Please also run it on the HPC cluster, just to make sure).
• Use the peak performance rather than the average performance from benchmarking results.
• In error calculation, report average and 95% confidence interval.
• Compute and report the prediction error for the Matrix Multiplication program ( third bullet in PA14 below, posted on Nov 28).
• Report the error of prediction with points selection (second bullet in PA14 below, posted on Nov 28), for the "mixed" I/O configuration (16 data points) and the MM computation (14 data points).

Project Assignment 14

Attention: all slides and drafts due on noon, Wed, Dec 3rd.

Continue prediction evaluation

• Improve the error evaluation by tossing out known points in the average error calculation.
• For existing MPI-IO prediction with 16 data points, try to select 2, 3, 4, and 5 evenly spread points in the target curve as known points. Perform the prediction and compare the error with the average error from the random-point predictions (with corresponding number of known points).

Extend experiments to computational modules

• Benchmark and perform predictions using a parallel matrix multiplication code. When doing the prediction, fix the number of processors (say 16) and the array distribution style (say BLOCK style). The purpose is to simulate simply feeding matrices to a black-box matrix computation library. The x-axis are formed by input matrix sizes and shapes:
  C = A * B, where C is m*n, A is m*k and B is k*n.
  Let s be a relatively large number (depending on how large the matrices will be, for example, 8096). Benchmark the following 14 input size combinations on both clusters:
  

  m	n	k
  s	s	s
  s	s	1
  s	1	s
  1	s	s
  1	s	1
  s	1	1
  1	1	s
  s	s	s/2
  s/2	s	s
  s	s/2	s
  s/2	s	s/2
  s/2	s/2	s
  s	s/2	s/2
  s/2	s/2	s/2

Draft the term project report.

• Write a preliminary version of the term report using the given latex template. Turn in a PDF or PS file.