phd-thesis/manuscrit/60_staticdeps/99_conclusion.tex

30 lines
1.7 KiB
TeX
Raw Normal View History

2024-05-05 20:11:53 +02:00
\section*{Conclusion}
In this chapter, we studied data dependencies within assembly kernels; and more
specifically, data dependencies occurring through memory accesses, which we
call \emph{memory-carried dependencies}. \cesasme{}'s analysis showed in
\autoref{chap:CesASMe} that this kind of dependency was responsible for a
significant portion of state-of-the-art analyzers' prediction errors.
We introduce \staticdeps{}, a heuristic approach based on random values as
representatives of abstract values. This approach is able to find data
dependencies, including memory-carried ones, loop-carried or not, leveraging
semantics of the assembly code provided by \valgrind{}'s \vex. It is, however,
still unable to find aliasing addresses whose source of aliasing is outside of
the studied block's scope ---~and, as such, suffers from the \emph{lack of
context} pointed out in the previous chapter.
\medskip{}
Our evaluation of \staticdeps{} against a dynamic analysis baseline,
\depsim{}, shows that it only finds about 60\,\% of the existing dependencies.
We however enrich \uica{} with \staticdeps{}, and find that it performs on the
full \cesasme{}'s dataset as well as \uica{} alone on the pruned dataset of
\cesasme{}, removing memory-carried bottlenecks. From this, we conclude that
\staticdeps{} is very successful at finding the data dependencies through
memory that actually matter from a performance analysis perspective. We also
find that, despite being written in pure Python, \staticdeps{} is at least
30$\times$ faster than its C dynamic counterpart, \depsim; as such, we expect
a compiled and optimized implementation of \staticdeps{} to be two to three
orders of magnitude faster than \depsim{}.