148 lines
6.5 KiB
TeX
148 lines
6.5 KiB
TeX
%% SCRAP: papers/EXECUTIVE_SUMMARY
|
|
%% SOURCE: docs/working/papers/EXECUTIVE_SUMMARY.md
|
|
%% STATUS: CURRENT
|
|
%% FITS: ssrn/intro, vol3-research/intro
|
|
%% EDITORIAL: lifted — prose rewritten to press voice
|
|
|
|
\section{Executive Summary}
|
|
|
|
\subsection{The Problem}
|
|
|
|
Modern runtime systems face a persistent conflict between two design goals.
|
|
Adaptive systems---contemporary web browsers, mobile application runtimes,
|
|
JIT-compiled virtual machines---learn which code paths run most often and
|
|
reorganize themselves to accelerate those paths. They improve over time, but
|
|
their internal decisions vary from run to run; two executions of the same
|
|
program on the same machine may produce different optimization states. This
|
|
non-determinism is acceptable for web browsers but disqualifying for
|
|
safety-critical software, financial systems requiring auditable behavior, or
|
|
real-time systems with timing guarantees.
|
|
|
|
Deterministic systems---firmware for medical devices, aircraft flight
|
|
controllers, certified real-time kernels---run the same way on every
|
|
invocation. They can be formally verified, exhaustively tested, and
|
|
certified by standards bodies. But they cannot adapt: the optimization
|
|
configuration frozen at compile time may be far from optimal for the actual
|
|
workload seen at runtime.
|
|
|
|
StarForth resolves this conflict. The system is simultaneously adaptive
|
|
and completely predictable at the level of its algorithmic decisions.
|
|
|
|
\subsection{What Was Built}
|
|
|
|
StarForth is a FORTH-79 compliant virtual machine with a physics-driven
|
|
adaptive runtime. Using execution frequency as a proxy for thermal energy
|
|
(a deliberate conceptual metaphor, not a physics claim), the system tracks
|
|
how often each dictionary entry executes, applies exponential decay to reduce
|
|
the weight of stale executions, and maintains a small cache of the most
|
|
frequently executed words for accelerated lookup. Every decision in this
|
|
pipeline is deterministic: given the same execution sequence, the system
|
|
produces the same cache configuration on every run.
|
|
|
|
Seven feedback loops coordinate the adaptive behavior:
|
|
|
|
\begin{itemize}
|
|
\item Execution heat tracking and hot-words caching (Loop~1)
|
|
\item Rolling window of execution history (Loop~2)
|
|
\item Linear decay of quiescent entries (Loop~3)
|
|
\item Word-to-word transition prediction (Loop~4)
|
|
\item Variance-based window width inference via Levene's test (Loop~5)
|
|
\item Decay slope inference via exponential regression (Loop~6)
|
|
\item Adaptive heartbeat coordination (Loop~7)
|
|
\end{itemize}
|
|
|
|
All loops are independently togglable and formally characterized. Fixed-point
|
|
arithmetic (\Qtype\ throughout) eliminates IEEE-754 non-determinism across
|
|
architectures.
|
|
|
|
\subsection{Experimental Results}
|
|
|
|
A three-configuration design of experiments ran 90 trials (30 runs per
|
|
configuration):
|
|
|
|
\begin{itemize}
|
|
\item \textbf{C\_NONE} (baseline): all adaptive loops disabled
|
|
\item \textbf{C\_CACHE}: hot-words cache enabled, inference disabled
|
|
\item \textbf{C\_FULL}: all seven loops active
|
|
\end{itemize}
|
|
|
|
\paragraph{Result 1: algorithmic determinism.}
|
|
Across all 30 runs of C\_FULL, the cache hit rate was 17.39\% with a
|
|
coefficient of variation (CV) of exactly 0.00\%. The F-test for variance
|
|
homogeneity yields $F(29,29) \to \infty$, $p < 10^{-30}$. Even under
|
|
intentional thermal stress (sustained CPU load), the algorithmic decisions
|
|
did not change; only wall-clock runtime increased.
|
|
|
|
\paragraph{Result 2: adaptive convergence.}
|
|
C\_FULL improved from a mean of 10.20\,ms in early runs (1--15) to 7.61\,ms
|
|
in late runs (16--30), a statistically significant 25.4\% improvement
|
|
($t = 4.23$, $p < 0.001$, Cohen's $d \approx 5.08$). The baseline
|
|
configuration (C\_NONE) showed slight degradation over the same window;
|
|
C\_CACHE showed no meaningful improvement. Only the fully adaptive
|
|
configuration converged.
|
|
|
|
\paragraph{Result 3: variance decomposition.}
|
|
Runtime exhibits 60--70\% CV due to OS scheduler noise, thermal variation,
|
|
and cache-line effects. Cache decisions exhibit 0.00\% CV. The two components
|
|
are statistically independent (Pearson $r = 0.03$, $p = 0.87$). The
|
|
adaptive algorithm is perfectly stable; the environment adds noise on top
|
|
of a deterministic foundation.
|
|
|
|
\subsection{Implications}
|
|
|
|
\paragraph{Verifiable adaptive systems.}
|
|
A system whose adaptive decisions are deterministic can, in principle, be
|
|
formally verified. The optimization state reachable from a given workload is
|
|
predictable, auditable, and reproducible---properties required for
|
|
safety-critical certification.
|
|
|
|
\paragraph{Reproducible performance characterization.}
|
|
Benchmark results become portable: different researchers running the same
|
|
workload reach the same adaptive steady state, enabling apples-to-apples
|
|
comparison of optimization strategies.
|
|
|
|
\paragraph{Adaptation without non-determinism.}
|
|
The results demonstrate that statistical inference can guide optimization
|
|
without introducing randomness. The system adapts through arithmetic and
|
|
counting, not machine learning or stochastic search.
|
|
|
|
\subsection{Scope and Limitations}
|
|
|
|
This work does not claim that StarForth is the fastest virtual machine, that
|
|
its approach is superior to JIT compilation for all use cases, or that the
|
|
thermodynamic framework is a physical theory. The 0.00\% algorithmic CV
|
|
applies to cache decisions, not to total system runtime. The system is
|
|
validated on CPU-bound, deterministic FORTH programs; I/O-bound workloads,
|
|
programs with random control flow, and very short processes fall outside the
|
|
validated scope. Fifteen failure modes are documented in the companion
|
|
negative-results section.
|
|
|
|
\subsection{Reproducibility}
|
|
|
|
Full source code, raw experimental data from all 90 runs, and a
|
|
step-by-step reproduction protocol are publicly available. A Docker container
|
|
provides bit-for-bit exact reproduction against a pinned environment. The
|
|
authoritative reproduction command is:
|
|
|
|
\begin{lstlisting}[language=bash]
|
|
git clone https://github.com/rajames440/StarForth.git
|
|
cd StarForth
|
|
make fastest
|
|
./build/amd64/fastest/starforth --doe
|
|
\end{lstlisting}
|
|
|
|
Deviations from the expected 0.00\% algorithmic CV should be reported as bugs.
|
|
|
|
\subsection{Summary}
|
|
|
|
\begin{itemize}
|
|
\item An adaptive virtual machine can achieve 0.00\% algorithmic variance
|
|
across 90 experimental runs.
|
|
\item Adaptation and determinism are not mutually exclusive; statistical
|
|
inference drives optimization without introducing non-determinism.
|
|
\item Performance improves measurably (25.4\%) as the system converges to
|
|
steady state.
|
|
\item All data, code, and reproduction instructions are publicly available
|
|
for independent verification.
|
|
\end{itemize}
|