Downloading, Compiling and Running this Benchmark

• Downloading
• Compiling
• Running

Downloading

Requirements

We need next requirements for compiling and running R^2mark.

• C++ compiler
  We tested below compilers.
  • g++ 2.7.2
  • egcs-2.90.29
  • egcs-2.95.2
• Java compiler : JDK1.2
• MPI library
  • mpich-1.1.2
  • mpich-1.2.0
• mpiJava + MAEDA-patch

You need C++ compiler (we use g++ 2.7.2 or egcs-2.90.29 980515 (egcs-1.0.3 release))

• R^2mark Source code with Dataset (7.2MB) (including patched mpiJava-1.2)

• ---

• mpich : An C implementation of MPI (The Message Passing Interface standard)
• mpiJava web page
  

  `mpiJava 1.2beta' has some memory leak bugs, R^2mark Source code include fixed mpiJava 1.2beta. We informed that to mpiJava developer team, so if you can use next mpiJava version, use it. However, if that is not available, use the mpiJava-1.2 included in R^2mark Source code.

Compiling R^2mark

Please set `mpirun' command of the mpich in your path for compiling parallel version of R2mark. The `configure' program searches `mpirun' command and may find headers and libraries for MPI. If you did not set this, configure may ask it to you. Anyway, if you do not need the parallel version, you do not need anything.

1. Unpack the tar.gz file. This is common procedure of C++ and Java version. Once you do this C++ version and Java version source code is unpacked.
   

   gunzip R2mark-[release number].tar.gz
   tar xvf R2mark-[release number].tar
          

2. Compile :
   

   We define the directory R2mark-[release number] as ${R2markHOME}.

   cd ${R2markHOME}
   make r2mark
          

3. If you want to a sub package only,
   1. C++ version compiling : If you do not need C++ version, goto Java version compiling

      cd ${R2markHOME}/C++
      ./configure;
      make
      	      

   2. Java version compiling First, you have to build mpiJava. and build the Java version of R2Mark.

      cd ${R2MARKHOME}/Java/jp/ac/tohoku/is/archi/mpi2;
      ./configure
      make
      	      

4. Optional Test : In this optional test, created rendering images are not so good, since these are just for tests.
   
   • Minimal Test. Serial version.(test 1)
     • C++

       cd ${R2markHOME}/C++/sampledat/1PE/4tri;
       ./tes.sh
       		     

     • Java

       cd ${R2markHOME}/Java/jp/ac/tohoku/is/archi/mpi2/sampledat/1PE/4tri;
       chmod u+x ./tes.sh
       ./tes.sh
       		     

     You can find a save.test.out file in this directory. This is ppm file of rendered image. test 1 result: You can see a Gouraud shaded square.

   • Small Test. Serial version. (test 2)

     cd ${R2markHOME}/C++/sampledat/1PE/room16; (C++)
     cd ${R2markHOME}/Java/jp/ac/tohoku/is/archi/mpi2/sampledat/1PE/room16; (Java)
     ./tes.sh
     	      

     test 2 result: You can see a notebook computer in a room.

   • Parallel Test. (test 3)

     cd ${R2markHOME}/C++/sampledat/mpi; (C++)
     cd ${R2markHOME}/Java/jp/ac/tohoku/is/archi/mpi2/sampledat/parallel; (Java)
     
     ./tes.sh
     	      

     First, this script creates a `machines' file. This includes a list of rshable machine list. And calls `mpirun'.

     When you can use the ps command or the top command, you can see the three processes named `mpi2'. One is `host' and other two processes processing Radiosity and Ray-tracing in parallel (or concurrently).

     test 3 result: You can see a notebook computer in a room. This is same scene of test 2.

Sample result

For executing optional test, it takes time abound as below.

1. Minimal Test. Serial version.(test 1)
   • C++ : under one second
   • Java : around 20 seconds
2. Small Test. Serial version. (test 2)
   • C++ : around 15 seconds
   • Java : around 9 minutes
3. Parallel Test. (test 3)
   • C++ : around 12 minutes
   • Java : around 35 minutes

The test environment is

• Sony Vaio PCG-XR7S: PentiumIII 600 MHz
• Linux 2.2.14
• gcc egcs-2.91.66
• java version "1.2.2" Classic VM (build Linux_JDK_1.2_2_RC4, native threads, sunwjit)
  

Parallel result is very bad relative to serial one on the same scene, because this test is executed on one processor machine. If you run this program on one processor machine, task switch overhead and cache replacement overhead is very high. In Test 3, CPU time is only one minute.

Running

This benchmark includes some scripts for benchmarking. These scripts are under ${R2markHOME}/benchscript. The directory structure is next.

${R2markHOME} --+- benchscript ----+ parallel ----+ C++  ---+---- conf
                                   |              |         +---- office
                                   |              |         +---- soda
                                   |              |
                                   |              + Java ---+---- conf
                                   |                        +---- office
                                   |                        +---- soda
                                   + serial   ----+ C++ 
                                                  |
                                                  + Java

I prepare,

• Three scenes for rendering.
  • office ... the number of patches is about 100, 000, relatively vacancy scene.
  • soda ... the number of patches is about 130, 000, general scene.
  • conf ... the number of patches is about 220,000, complex and general scene.
• Three accuracy of lighting simulation. (What is accuracy?)
  • test ... Just only test. Energy residual convergence value is very low and sampling resolution is also low.
  • inaccurate ... Energy residual convergence value is from 40% to 70%. Sampling resolution is middle.
  • accurate ... energy residual convergence value is more than 90%. File sampling resolution. This is practical rendering.
• Two program for each language.
  • Serial
    No host process is needed. However, of course, the degree of parallelism equals one.
  • Parallel
    Even only one process calculates Radioisty and Ray-Tracing, host process is needed for controlling the whole program. In this time, the degree of parallelism is 1, 2, 4 and 8.

Most easy way to run the benchmarks.

• serial version
  

  cd ${R2markHOME}/benchscript/serial;
  ./run_benchmarks.sh
         

• parallel version

  cd ${R2markHOME}/benchscript/parallel;
  ./run_benchmarks.sh
         

This benchmark takes a long time to run. At Sun UltraII330Mhz, scene conf with accuracy under C++ version takes more than 1 day. (However, scene office with test takes only a few minutes.)