papers/lux-tee-computing-mesh/sections/08-performance-analysis.tex

\section{Performance Analysis}


\subsection{Throughput Scaling}

\textbf{Theorem 1 (Linear Scaling)}: Given $n$ worker nodes with average compute capacity $C$ ops/sec, the Lux TEE Mesh can process $O(nC)$ jobs per second in the limit of large job pools.

\textit{Proof sketch}: The gossip network has $O(\log n)$ latency for message propagation \cite{zap-protocol}. Job selection is decentralized, so no bottleneck exists. Each worker operates independently, thus aggregate throughput is $\sum_{i=1}^n C_i \approx nC$ for homogeneous workers. \qed

\subsection{Attestation Verification Latency}

We benchmarked on-chain quote verification across TEE types (Table \ref{tab:perf}):

\begin{table}[h]
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\textbf{TEE Type} & \textbf{Quote Size} & \textbf{Gas Cost} & \textbf{Latency (ms)} \\
\midrule
Intel SGX v4 & 432 bytes & 185,000 & 42 \\
Intel TDX & 1,024 bytes & 320,000 & 78 \\
AMD SEV-SNP & 1,184 bytes & 340,000 & 85 \\
NVIDIA GPU & 512 bytes & 210,000 & 51 \\
\bottomrule
\end{tabular}
\caption{On-chain attestation verification performance (Lux C-Chain, 8 gas/ms)}
\label{tab:perf}
\end{table}

\subsection{End-to-End Job Latency}

Median latency breakdown for a 1-second compute job:

\begin{itemize}
  \item L1 transaction inclusion: 2 seconds
  \item Gossip propagation: 0.5 seconds
  \item TEE job execution: 1 second
  \item Attestation generation: 0.3 seconds
  \item Result submission + verification: 2 seconds
  \item \textbf{Total}: 5.8 seconds
\end{itemize}

For long-running jobs (e.g., LLM inference $>$ 30 seconds), overhead is negligible ($<$ 20\%).