tupper (no@thanks.com) on 1/25/12 wrote:
>
>Here is a research work done by UMich and ARM. How much sense does it make for the conclusion :

So I absolutely believe in the "tons of I$, ITLB and branch
prediction misses" thing.

Having done performance analysis of Spec and real desktop
applications (about ten years ago, so that's probably not
all that far from a smartphone), I can tell that Spec is
a set of incredibly well-behaved applications in general.

That includes gcc, which I obviously think is the only app
in the whole Spec suite that is worth even quoting numbers
for - the other ones are too effing broken, and much too
amenable to various trickery.

Graphical applications end up doing several layers of
indirection through libraries - something that you never
see in Spec. In fact, most Spec runs are outright lying,
in that they tend to flatten out even the one layer of
libraries there is - inlining a ton, and usually statically
linking the rest.

So spec is orders of magnitude better than the crazy toy
benchmarks the embedded world tends to historically use,
but spec isn't nearly gnarly enough to show what goes on
in a desktop app.

The paper you point to seems quite interesting from a quick
look. And it's really telling how the spec benchmarks show
basically no I$ component at all, but the interactive
ones do.

That said, there is one exception, and that one is not
shown in the paper. High-performance native games tend to
behave more like "traditional" loads than desktop loads.
Lots more tight loops etc. Of course, the actual graphics
stack is often quite big, and the 3D driver can get very
hard to gather data from.

Linus