papers/zoo-consensus/sections/08-empirical-evaluation.tex

\section{Empirical Evaluation}


\subsection{Testnet Configuration}

Evaluation conducted on Zoo Network testnet (Q1 2026):
\begin{itemize}
\item \textbf{Validators}: 100, 500, 1000, 2000
\item \textbf{Hardware}: NVIDIA A100 40\,GB (consensus), AMD EPYC 7763 (CPU baseline)
\item \textbf{Network}: 100\,Mbps inter-validator links, geographically distributed (US, EU, APAC)
\item \textbf{Block size}: 1\,MB, 5\,MB, 10\,MB
\end{itemize}

\subsection{Finality Latency}

\begin{table}[h]
\centering
\begin{tabular}{lcccc}
\toprule
\textbf{Validators} & \textbf{BLS only (ms)} & \textbf{Corona only (ms)} & \textbf{Dual-path (ms)} & \textbf{Overhead} \\
\midrule
100 & 210 & 380 & 390 & 2.6\% \\
500 & 340 & 620 & 640 & 3.2\% \\
1000 & 480 & 890 & 920 & 3.4\% \\
2000 & 710 & 1,340 & 1,380 & 3.0\% \\
\bottomrule
\end{tabular}
\caption{Median finality latency with GPU batch verification}
\label{tab:finality}
\end{table}

The dual-path overhead is consistently $< 4\%$ over \Corona{}-only, demonstrating that parallel execution effectively hides the BLS path latency behind the slower \Corona{} path.

\subsection{Throughput}

With 1,000 validators and 5\,MB blocks, Zoo achieves:
\begin{itemize}
\item \textbf{Block production}: 2 blocks/second
\item \textbf{Transaction throughput}: 4,200 TPS (average transaction size 1.2\,KB)
\item \textbf{Finality throughput}: 1.1 finalized blocks/second (920\,ms finality)
\end{itemize}