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.
72 lines
2.8 KiB
TeX
72 lines
2.8 KiB
TeX
\section{Comparison with Centralized Clouds}
|
|
\label{sec:comparison}
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\caption{Sovereignty comparison between Lux provider model and centralized cloud platforms.}
|
|
\label{tab:comparison}
|
|
\begin{tabular}{p{3.5cm}ccccc}
|
|
\toprule
|
|
\textbf{Property} & \textbf{Lux} & \textbf{AWS} & \textbf{GCP} & \textbf{Vercel} & \textbf{Fly.io} \\
|
|
\midrule
|
|
User holds encryption keys & Yes & No & No & No & No \\
|
|
Provider-independent data & Yes & No & No & No & Partial \\
|
|
On-chain SLA enforcement & Yes & No & No & No & No \\
|
|
Switch without data loss & Yes & No & No & No & Partial \\
|
|
Switch without code rewrite & Yes & No & No & No & Partial \\
|
|
Verifiable provider quality & Yes & No & No & No & No \\
|
|
Compute privacy (TEE/FHE) & Yes & Partial & Partial & No & No \\
|
|
Open market pricing & Yes & No & No & No & No \\
|
|
Censorship resistance & Yes & No & No & No & No \\
|
|
Decentralized availability & Yes & No & No & No & Partial \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
\subsection{Performance Comparison}
|
|
|
|
The decentralized model introduces overhead from encryption, chain
|
|
receipts, and market clearing. We quantify these costs:
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\caption{Performance overhead of the Lux provider model relative to direct cloud access.}
|
|
\label{tab:performance}
|
|
\begin{tabular}{lrrr}
|
|
\toprule
|
|
\textbf{Operation} & \textbf{Centralized} & \textbf{Lux (TEE)} & \textbf{Lux (FHE)} \\
|
|
\midrule
|
|
Object read (64 KB) & 2 ms & 3 ms & N/A \\
|
|
Object write (64 KB) & 5 ms & 8 ms & N/A \\
|
|
Inference (7B model, 100 tokens) & 1.2 s & 1.4 s & 4.8 s \\
|
|
Receipt anchoring & N/A & 50 ms & 50 ms \\
|
|
Provider switch (1 TB) & 12--48 h & 2--6 h & 2--6 h \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
The overhead is modest for TEE mode (15--60\% depending on operation)
|
|
and significant for FHE mode ($3\times$--$4\times$ for compute). The
|
|
receipt anchoring cost (50~ms amortized over batched receipts) is
|
|
negligible for most applications. Provider switch time is substantially
|
|
faster than centralized migration because no re-encryption is required
|
|
and parallel transfer from multiple replicas is supported.
|
|
|
|
\subsection{Lock-In Analysis}
|
|
|
|
We formalize the switching cost for each platform:
|
|
|
|
\begin{definition}[Switching Cost]
|
|
The switching cost $\mathsf{SC}(\mathcal{P}_1, \mathcal{P}_2)$ from provider
|
|
$\mathcal{P}_1$ to $\mathcal{P}_2$ is the total cost (in time, engineering
|
|
effort, and data risk) of migrating an application with $D$ bytes of data,
|
|
$A$ API surface area, and $I$ infrastructure dependencies.
|
|
\end{definition}
|
|
|
|
For centralized clouds, $\mathsf{SC}$ is dominated by API incompatibility
|
|
(rewriting application code to use $\mathcal{P}_2$'s proprietary APIs) and
|
|
data re-encryption. For the Lux model, $\mathsf{SC}$ is dominated by data
|
|
transfer time, which is $O(D / \mathsf{bandwidth})$ with no API rewrite
|
|
required (providers implement the same open protocol).
|
|
|