• AIM7Benchmark

The AIM 7 Benchmark

AIM7 is a multiuser job-throughput benchmark, which attempts to model a development environment; multiple users doing lots of things (viewing files, running troff, cc, editors etc). We are using the ReAIM implementation. Some further details can be found in the IA64 Wiki.

Running on K42

NB: until this patch is merged into K42, you will also need to apply it to your tree before beginning.

1. Download reaim from http://re-aim-7.sourceforge.net/re-aim-7.tar.gz and unpack it

2. Apply the following patch to src/pipe_test.c. This patch stops reaim from using UNIX-domain datagram sockets (currently unimplemented), and fixes a couple of other things that caused it to break on K42. Obviously this means that you can't directly compare results from this modified benchmark to other reaim output.

   --- re-aim-7.orig/src/pipe_test.c       2003-11-06 08:50:31.000000000 +1100
   +++ re-aim-7/src/pipe_test.c    2006-03-29 12:36:11.775366920 +1100
   @@ -24,6 +24,8 @@
    ****************************************************************/
    /* increased sizes in shared memory - re-aim */
    #define _INCLUDE_HPUX_SOURCE   /* does not compile & link on HP w/o this */
   +#define NO_SOCKETPAIR /* AB hack for K42 */
   +
    #include <stdio.h>
    #include <signal.h>
    #include <errno.h>
   @@ -152,7 +154,7 @@
    #else
           register_test("stream_pipe", 1, "100", stream_pipe, 100,
                         "Stream Pipe Messages");
   -       register_test("dgram_pipe", 1, "100", dgram_pipe, 100,
   +       register_test("dgram_pipe", 1, "100", udp_test, 100,
                         "DataGram Pipe Messages");
    #endif
           register_test("pipe_cpy", 1, "100", pipe_cpy, 100,
   @@ -378,7 +380,7 @@
           status = rd = open(fn1, O_RDONLY | O_NDELAY);   /* open the fifo to read */
           if (status < 0)
                   FATAL("Cannot open fifo readonly");     /* if error, stop */
   -       status = wr = open(fn1, O_RDWR | O_NDELAY);     /* open it again to write */
   +       status = wr = open(fn1, O_WRONLY | O_NDELAY);   /* open it again to write */
           if (status < 0)
                   FATAL("Cannot open fifo writeonly");    /* if error, stop */
           status = read_write_close(i64, "Named Pipe", rd, wr);   /* do the test */
   @@ -495,7 +497,7 @@
                   errno = 0;      /* clear errors */
                   result = read(fd, buf, total);  /* read some */
                   if (result < 0) {
   -                       if (errno == EINTR)
   +                       if (errno == EINTR || errno == EAGAIN)
                                   continue;       /* try again if interrupted */
                           return result;  /* return errors here */
                   }
   @@ -521,7 +523,7 @@
                   errno = 0;      /* clear error number */
                   result = write(fd, buf, total); /* write some */
                   if (result < 0) {       /* handle unusual case */
   -                       if (errno == EINTR)
   +                       if (errno == EINTR || errno == EAGAIN)
                                   continue;       /* if interrupted, loop */
                           return result;  /* return errors here */
                   }
   @@ -609,8 +611,8 @@
           }
           if (shmid != -1) {      /* only clear if set */
                   if (shmctl(shmid, IPC_RMID, 0) < 0)     /* free the shared memory */
   -                       fprintf(stderr,
   -                               "clear_ipc(): can't remove shared memory\n");
   +                       (void)0; /*fprintf(stderr,
   +                               "clear_ipc(): can't remove shared memory\n");*/
                   shmid = -1;
           }
           if (sem != -1) {        /* only clear if set */
   

3. Run ./configure --host=powerpc64-linux, and then make, to cross-compile the benchmark

4. Install the final result somewhere. You'll also have to manually copy the reaim.config file and the entire scripts subdirectory to $prefix/share/reaim/. At the end of it, you should have a directory tree that looks like this:

   bin/reaim
   share/reaim/fakeh.tgz
   share/reaim/reaim.config
   share/reaim/workfile.aim9
   share/reaim/workfile.all_utime
   share/reaim/workfile.alltests
   share/reaim/workfile.compute
   share/reaim/workfile.custom
   share/reaim/workfile.dbase
   share/reaim/workfile.disk
   share/reaim/workfile.five_sec
   share/reaim/workfile.fserver
   share/reaim/workfile.high_systime
   share/reaim/workfile.long
   share/reaim/workfile.new_dbase
   share/reaim/workfile.new_fserver
   share/reaim/workfile.shared
   share/reaim/workfile.short
   share/reaim/scripts/aim9-compare.pl
   share/reaim/scripts/aim_1.sh
   share/reaim/scripts/aim_2.sh
   share/reaim/scripts/aim_3.sh
   share/reaim/scripts/example_load.sh
   share/reaim/scripts/instruct.sh
   share/reaim/scripts/run.10.users.sh
   share/reaim/scripts/runall.sh
   share/reaim/scripts/setup.sh
   share/reaim/scripts/true
   share/reaim/scripts/xover_load.sh
   

5. Define REAIM_PKG to the location of your installed reaim tree (eg. in Make.paths or MakePaths.proto for your site).

6. Rebuild and boot K42, then login and run /home/reaim/runReAIM. Use runReAIM -h to see the available options.

No doubt something will break. Feel free to ask AndrewBaumann or JeremyKerr for advice.

Results

NB: these are all old data, and it's not clear what workfile was in use (it probably wasn't workfile.alltests, which the current script defaults to). Poor K42 results in comparison to Linux are expected, because of the high cost of fork in K42.

Running on an xserve at UNSW

For a noDeb build of the current (2005-10-18) HEAD on a 2-CPU G5 xserve, I'm getting around 39500 jobs per minute.

Running on k1 at watson

Issues

• The script doesn't leave the system in a terribly useful state, it seems it is tricky to undo the results of the bindFS call I use to get a chroot-style environment. It would be nice if someone implemented chroot.