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.
91 lines
3.4 KiB
TeX
91 lines
3.4 KiB
TeX
\section{Compute and Resource Markets}
|
|
\label{sec:markets}
|
|
|
|
The Lux provider model organizes compute, storage, and bandwidth as
|
|
continuous double-auction markets with on-chain settlement. Providers
|
|
post asks (offers to sell service at a given price). Users post bids
|
|
(willingness to pay for service at a given quality level). A matching
|
|
engine on the C-Chain clears the market each epoch.
|
|
|
|
\subsection{Market Structure}
|
|
|
|
Each resource type has an independent order book:
|
|
|
|
\begin{table}[H]
|
|
\centering
|
|
\caption{Resource market structure for the Lux decentralized cloud.}
|
|
\label{tab:markets}
|
|
\begin{tabular}{llll}
|
|
\toprule
|
|
\textbf{Market} & \textbf{Unit} & \textbf{Quality Dimensions} & \textbf{Settlement} \\
|
|
\midrule
|
|
Inference & GPU-second & Latency, throughput, model & Per-request \\
|
|
Storage & GB-month & Durability, availability, region & Per-epoch \\
|
|
Bandwidth & GB-transfer & Latency, throughput, region & Per-request \\
|
|
Recovery & Share-epoch & Availability, threshold & Per-epoch \\
|
|
Gateway & Request & Latency, concurrent connections & Per-request \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
\subsection{Inference Market}
|
|
|
|
The inference market is the most complex, as it must match heterogeneous
|
|
compute capabilities (GPU type, VRAM, supported models) with heterogeneous
|
|
demand (model size, batch size, latency requirements).
|
|
|
|
An inference ask has the form:
|
|
|
|
\begin{equation}
|
|
\mathsf{Ask}_{\text{infer}} = (\mathsf{gpu\_type}, \mathsf{vram}, \mathsf{models}[], \tau_{\max}, p_{\text{per\_token}})
|
|
\end{equation}
|
|
|
|
An inference bid has the form:
|
|
|
|
\begin{equation}
|
|
\mathsf{Bid}_{\text{infer}} = (\mathsf{model}, \mathsf{max\_tokens}, \tau_{\max}, p_{\max}, \mathsf{mode})
|
|
\end{equation}
|
|
|
|
where $\mathsf{mode} \in \{\mathsf{TEE}, \mathsf{FHE}, \mathsf{plaintext}\}$
|
|
specifies the privacy level. TEE and FHE modes command a premium (typically
|
|
$1.3\times$ and $3\times$ respectively) due to the computational overhead.
|
|
|
|
\subsection{Market Clearing}
|
|
|
|
The matching engine runs a modified second-price auction each epoch.
|
|
For each resource type, bids and asks are sorted by price. Matches
|
|
are made greedily by descending bid price against ascending ask price.
|
|
The clearing price for each match is the ask price (second-price
|
|
semantics), incentivizing truthful bidding:
|
|
|
|
\begin{equation}
|
|
p_{\text{clear}}(b, a) = p_a \quad \text{where } p_b \geq p_a
|
|
\end{equation}
|
|
|
|
\begin{theorem}[Truthful Bidding]
|
|
Under second-price semantics with on-chain settlement, truthful bidding
|
|
(reporting true valuation) is a weakly dominant strategy for users.
|
|
\end{theorem}
|
|
|
|
\begin{proof}
|
|
Standard Vickrey auction argument. If a user bids above true valuation
|
|
and wins, they may pay more than the resource is worth. If they bid below
|
|
and lose, they miss a profitable trade. Bidding true valuation maximizes
|
|
expected surplus in all cases.
|
|
\end{proof}
|
|
|
|
\subsection{Dynamic Pricing}
|
|
|
|
The base price for each resource type adjusts each epoch based on
|
|
utilization, following an EIP-1559-inspired mechanism:
|
|
|
|
\begin{equation}\label{eq:pricing}
|
|
p_{t+1}^{\text{base}} = p_t^{\text{base}} \cdot \left(1 + \frac{1}{8} \cdot \frac{U_t - U_{\text{target}}}{U_{\text{target}}}\right)
|
|
\end{equation}
|
|
|
|
where $U_t$ is the utilization ratio (fraction of available capacity
|
|
consumed) and $U_{\text{target}} = 0.5$ is the target utilization.
|
|
When utilization exceeds 50\%, prices rise; when it falls below, prices
|
|
decrease. The $1/8$ dampening factor prevents price oscillation.
|
|
|