[distcc] Benchmark of distcc, ccache, and distcc with ccache

[Dan Kegel]

We were curious how much ccache helped
relative to distcc, so we ran the distcc benchmarks
using various combinations of distcc and ccache.
We think we did everything right (e.g. we avoided using
NFS for DISTCC_DIR or CCACHE_DIR-- boy, was that slow),
but you can check for yourself - the data and scripts we used are online at
http://students.cs.tamu.edu/dzhang/distcc_result.html

Here's an excerpt of the results.
For the benefit of those whose mail clients use proportional
fonts, I'm sending this using HTML mail and picking a fixed
width font (though every fiber of my being hates doing so).

Runtime in seconds (smaller is better), average of five interleaved runs

                                    ccache+
bench           local    ccache    distcc-j8     distcc
                       cold   hot  cold   hot   -j4   -j8
gdb-5.3         54     101    38    39    26    41    32
httpd-2.0.43    78     86     43    62    44    55    55
samba-2.2.7     90     103    35    56    26    71    45
linux-2.5.51    261    259    86    77    51    106   66
wine-0.9.4      1335   1446   535   583   394   687   487

Same data, divided by first column so 1.0 is "same as local build"
Smaller is better

                                    ccache+
bench           local    ccache    distcc-j8     distcc
                       cold   hot  cold   hot   -j4   -j8
gdb-5.3         1.00   1.87   0.69  0.72  0.49  0.75  0.59
httpd-2.0.43    1.00   1.10   0.55  0.80  0.56  0.71  0.70
samba-2.2.7     1.00   1.14   0.39  0.62  0.29  0.79  0.50
linux-2.5.51    1.00   0.99   0.33  0.30  0.19  0.40  0.25
wine-0.9.4      1.00   1.08   0.40  0.44  0.29  0.51  0.36

The client was a uniprocessor 3.4GHz Pentium 4 with 3GB RAM.
The servers were eight dual 2.6GHz Pentium 4's with 2GB RAM.
Distcc was 2.18.3; ccache was 2.2.
Both client and servers were dedicated - nobody else was using the compile
cluster.
The network was a studly corporate LAN, but other folks were around,
so the results are slightly suspect; we ran the whole test five times,
and used the average times, to guard against that a bit.

No big surprises there, but I thought folks might like to see the data.
- Dan

--
Wine for Windows ISVs: http://kegel.com/wine/isv
-------------- next part --------------
HTML attachment scrubbed and removed