Reviewers: Craig Silverstein

ACE), and normalize syntax for existing ones.  I've tested that we can
at least build all of these projects locally, as long as all neccesary
dependent packages are installed.

Reviewed by fergus

server to terminate before exiting.

Reviewers: Nils Klarlund, Craig Silverstein

mode.

This is harmless I think, and it's easy to forget this option.  It is of
importance to at least Samba-2.2.7.

changed default C++ compiler to 'c++' instead of 'cxx'. 'c++' is the more common
name I believe.  Witht this change, both compilers must exist for benchmark.py
to run.

REVIEW: csilvers@google.com

This CL adds reporting of:

 - number of processors
 - total cpu time
 - cpu utilization
 - include server cpu time

I corrected issues with overly loose specification of compiler locations. (I
developed this CL under Cygwin and encountered new problems.)  This resulted in
evil recursive invocations of distcc.

I also:

-- removed an unnecessary parameter from Build(..) 

-- pruned the number of decimals reported from 4 to 1.

-- removed the dependency of 'Numeric', which is a non-standard package and
   rewrote the math do to std deviation

-- use shell built-in 'time' instead of Python time functionality

-- corrected deficient error reporting: builds would fail with non-zero status
   and not be reported as failing.

REVIEW: csilvers@google.com

TESTS:
    DISTCC_HOSTS="--randomize `/home/klarlund/svn-distcc/distcc/lsdistcc -P1 -k100 -d`" PATH=`pwd`:$PATH bench/benchmark.py --cc='/usr/crosstool/v10/gcc-4.2.1-glibc-2.3.2/x86_64-unknown-linux-gnu/bin/x86_64-unknown-linux-gnu-gcc'  -n 1 -c 'pump,h40,j120' -c 'dist,h40,j32' linux-2.6.25

Part of the printout for hello-2.1.1:

Local number of CPUs: 4
Linux klarlund-warp.nyc.corp.google.com 2.6.18.5-gg34workstation-mixed64-32 #1 SMP Thu May 8 01:31:23 UTC 2008 x86_64 GNU/Linux
project               compiler                             time       s.d.   CPU time   CPU util  incl serv
hello-2.1.1           pump_h39_j120                        0.9s       0.0s       0.5s      13.5%       0.2s
hello-2.1.1           dist_h39_j32                         0.4s       0.1s       0.3s      18.1%


(For this small program, pump is a disadvantage.)

unnecessarily.  Most prominently, no longer re-run 'tar xf &&
configure' every benchmark run.  Instead, if configure was
successfully executed last benchmark run, just do 'make clean'.  This
is equivalent to what we do now between build runs, when the
repeat-count is > 1.

This feature is not turned on by default, but must be enabled with the
new --force=0 flag.  --force=1 keeps the same behavior as before: we
always re-run configure in each build-directory.  --force=2 adds more
repeat-work than before; in --force=2 mode, we always re-download the
package from the web, even if we had successfully downloaded it
before.

Note that the 'tar' command currently warms the file-cache, which
makes build times more consistent, so there is so cost to setting
--force=0, in terms of the data collected.

Reviewed by klarlund

This makes the linux kernel build with pump-mode when the benchmark script is
used.

Also, include the '-t' option to the include server; that prints out the elapsed
and CPU times at the end of the build.

REVIEW: csilvers@google.com
TESTS: (on quad-cpu workstation)
      DISTCC_HOSTS="--randomize `/home/klarlund/svn-distcc/distcc/lsdistcc -P1 -k100 -d`" PATH=`pwd`:$PATH bench/benchmark.py --cc='/usr/crosstool/v10/gcc-4.2.1-glibc-2.3.2/x86_64-unknown-linux-gnu/bin/x86_64-unknown-linux-gnu-gcc' --cxx='void' -n 5 -c 'pump,h40,j120' -c 'dist,h40,j32' linux-2.6.25

project               compiler                             time      s.d.
linux-2.6.25          pump_h39_j120                    69.6525s    3.0969s
linux-2.6.25          dist_h39_j32                    103.8416s   10.8041s

(typically the only change is in the FSF street address).  Add Google
copyright line in some places it was missing.  Add GPLv2 notice to
avahi patches.

OKed by fergus