papers/hanzo-network-architecture/sections/10-performance-evaluation.tex

\section{Performance Evaluation}


\subsection{Experimental Setup}

Hardware:
\begin{itemize}
\item CPU: AMD EPYC 7763 (64 cores)
\item GPU: NVIDIA H100 80GB
\item Memory: 512GB DDR4
\item Storage: NVMe SSD (Gen4)
\item Network: 100Gbps
\end{itemize}

Software:
\begin{itemize}
\item OS: Ubuntu 22.04 LTS
\item Rust: 1.75.0
\item Docker: 24.0.7
\item Kubernetes: 1.28
\end{itemize}

\subsection{Consensus Throughput}

\begin{table}[h]
\centering
\begin{tabular}{|l|c|c|}
\hline
\textbf{Configuration} & \textbf{Blocks/sec} & \textbf{Finality} \\
\hline
Single node (local) & 9,200,000 & Immediate \\
3 nodes (LAN) & 45,000 & 2.1s \\
10 nodes (LAN) & 12,000 & 3.8s \\
50 nodes (WAN) & 3,500 & 8.2s \\
\hline
\end{tabular}
\caption{Consensus throughput under different topologies}
\label{tab:consensus}
\end{table}

\subsection{PQC Overhead}

\begin{table}[h]
\centering
\small
\begin{tabular}{|l|c|c|}
\hline
\textbf{Operation} & \textbf{Classical} & \textbf{PQC} \\
\hline
Key agreement & 0.12 ms & 0.13 ms \\
Signature gen & 0.08 ms & 0.24 ms \\
Signature verify & 0.09 ms & 0.12 ms \\
\hline
\end{tabular}
\caption{Classical (X25519/Ed25519) vs PQC latency}
\label{tab:pqc_overhead}
\end{table}

PQC adds negligible overhead ($<$0.2ms) compared to network latencies (typically 10-100ms).

\subsection{Runtime Performance}

\begin{table}[h]
\centering
\small
\begin{tabular}{|l|c|c|c|}
\hline
\textbf{Runtime} & \textbf{Cold Start} & \textbf{Memory} & \textbf{CPU} \\
\hline
Docker & 2.3s & 512MB & 0.5 cores \\
Kubernetes & 4.1s & 768MB & 0.6 cores \\
WASM & 8ms & 64MB & 0.1 cores \\
Firecracker & 124ms & 128MB & 0.3 cores \\
\hline
\end{tabular}
\caption{Runtime cold start and resource consumption}
\label{tab:runtime_perf}
\end{table}

\subsection{HLLM Routing Latency}

\begin{table}[h]
\centering
\begin{tabular}{|l|c|}
\hline
\textbf{Component} & \textbf{Latency} \\
\hline
Regime detection & 0.8 ms \\
Adapter retrieval & 0.3 ms \\
EFE calculation & 1.2 ms \\
Model selection & 0.5 ms \\
\textbf{Total routing overhead} & \textbf{2.8 ms} \\
\hline
\end{tabular}
\caption{HLLM routing latency breakdown}
\label{tab:hllm_latency}
\end{table}

\subsection{Embedding Performance}

\begin{table}[h]
\centering
\small
\begin{tabular}{|l|c|c|}
\hline
\textbf{Model} & \textbf{Throughput} & \textbf{Latency} \\
\hline
Qwen3-8B (GPU) & 1,200 emb/s & 8.3 ms \\
Qwen3-8B (CPU) & 45 emb/s & 222 ms \\
Qwen3-4B (GPU) & 2,100 emb/s & 4.8 ms \\
Ollama remote & 180 emb/s & 55 ms \\
\hline
\end{tabular}
\caption{Embedding generation performance (batch=32)}
\label{tab:embedding_perf}
\end{table}