mirror of
https://github.com/lux-apps/papers.git
synced 2026-10-30 18:15:21 +00:00
zeekay
abf99d05fc
Restructured from flat .tex/.pdf files into per-paper subdirectories. Added Antje Worring as co-author on all papers. Synced PDFs to pdfs/ for download site.
81 lines
2.8 KiB
TeX
81 lines
2.8 KiB
TeX
\section{Benchmark 3: Validator Scalability}
|
|
|
|
\label{sec:scalability}
|
|
%% ============================================================================
|
|
|
|
We measure the \emph{degradation factor}: throughput at $n=1{,}000$ divided by
|
|
throughput at $n=21$.
|
|
|
|
\begin{table}[h]
|
|
\centering
|
|
\begin{tabular}{@{}lrrl@{}}
|
|
\toprule
|
|
\textbf{Protocol} & \textbf{TPS @ 21} & \textbf{TPS @ 1000} & \textbf{Retention} \\
|
|
\midrule
|
|
Snow & 5{,}200 & 3{,}900 & 75.0\% \\
|
|
\textbf{Quasar} & 18{,}200 & 4{,}800 & \textbf{26.4\%} \\
|
|
Narwhal+Tusk & 46{,}000 & 12{,}100 & 26.3\% \\
|
|
HotStuff & 8{,}500 & 1{,}400 & 16.5\% \\
|
|
Tendermint & 4{,}100 & 180 & 4.4\% \\
|
|
Casper FFG & 30 & 30 & 100\%$^*$ \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Throughput retention from $n=21$ to $n=1{,}000$. $^*$Casper FFG throughput is
|
|
execution-bound, not consensus-bound; the value is misleading.}
|
|
\label{tab:scalability}
|
|
\end{table}
|
|
|
|
\begin{figure}[h]
|
|
\centering
|
|
\begin{tikzpicture}
|
|
\begin{axis}[
|
|
width=0.95\textwidth,
|
|
height=7cm,
|
|
xlabel={Number of validators ($n$)},
|
|
ylabel={Throughput retention (\%)},
|
|
xmin=0, xmax=1100,
|
|
ymin=0, ymax=110,
|
|
legend style={at={(0.98,0.98)}, anchor=north east, font=\small},
|
|
grid=major,
|
|
grid style={gray!20},
|
|
]
|
|
% Quasar
|
|
\addplot[luxpurple, very thick, mark=square*, mark size=3pt] coordinates {
|
|
(21, 100) (100, 57.1) (500, 35.7) (1000, 26.4)
|
|
};
|
|
% Narwhal+Tusk
|
|
\addplot[narworange, thick, mark=triangle*, mark size=3pt] coordinates {
|
|
(21, 100) (100, 83.7) (500, 41.3) (1000, 26.3)
|
|
};
|
|
% Snow
|
|
\addplot[avaxred, thick, mark=diamond*, mark size=3pt] coordinates {
|
|
(21, 100) (100, 92.3) (500, 82.7) (1000, 75.0)
|
|
};
|
|
% Tendermint
|
|
\addplot[cosgray, thick, mark=o, mark size=3pt] coordinates {
|
|
(21, 100) (100, 29.3) (500, 9.8) (1000, 4.4)
|
|
};
|
|
% HotStuff
|
|
\addplot[hotred, thick, mark=pentagon*, mark size=3pt] coordinates {
|
|
(21, 100) (100, 60.0) (500, 28.2) (1000, 16.5)
|
|
};
|
|
\legend{Quasar, Narwhal+Tusk, Snow, Tendermint, HotStuff}
|
|
\end{axis}
|
|
\end{tikzpicture}
|
|
\caption{Normalized throughput retention as validator count increases. Snow's sampling-based
|
|
approach degrades most gracefully. Tendermint exhibits cliff-edge degradation.}
|
|
\label{fig:scalability}
|
|
\end{figure}
|
|
|
|
\subsection{Analysis}
|
|
|
|
Snow's sampling approach ($K=25$ regardless of $n$) provides the flattest scalability
|
|
curve---each validator only communicates with a fixed-size sample. The trade-off is
|
|
probabilistic (not deterministic) finality and higher absolute latency.
|
|
|
|
Quasar matches Narwhal+Tusk in retention (26\%) while providing 2.3$\times$ faster
|
|
finality. Tendermint's O($n^2$) all-to-all communication produces catastrophic
|
|
degradation: a 95.6\% throughput loss from 21 to 1{,}000 validators.
|
|
|
|
%% ============================================================================
|