Wanda, "That's a fine explanation for added performance, but it doesn't account for 7x."
If you recall, the whole 179.art fits into US-III L2
cache. If for some stupid reason some portions of old code
used malloc() back and forth, and newer versions
of compiler eliminate this, 10x and more is possible.
For much smaller caches perhaps no one was doing this
check, but for huge 8MB they might start looking into
it. Alpha is doing well on SPEC mostly because of its
huge caches.
Being caught in benchmark cheating is a big risk for
a big company. There were gradual improvements in this
subtest over subsequent versions of their compiler,
and there is more than one submission with this "anomaly".
It is easy to call foul especially by those who has
no profiling, no tracing, and no performance counter
data, and in general little clue about these matters.
I guess the US-III hardware simply overgrew the
179.art test challenges, as it has happened before with
SPEC92, SPEC95, and how it is happening with SPEC2k.
- Ali |
