mirror of
https://github.com/lux-apps/papers.git
synced 2026-10-30 18:15:21 +00:00
zooqueen
a1f77cacc9
Ringtail (Boschini et al., ePrint 2024/1113) is third-party prior art; Lux's Corona (Ring-LWE = Module-LWE rank-1) builds on it; Pulsar is Module-LWE rank-k. Retired the conflated boschini2024corona/corona2024/coronago/... bib keys to the single canonical ringtail/corona/pulsar/magnetar keys, rewrote \cite + \bibitem by context, and fixed 'Pulsar (Ring-LWE)' mislabels (Corona where corona.metal/ single-ring/Ringtail-derived; Pulsar where Module-LWE rank-k/ML-DSA). Cryptographer-led.
122 lines
4.2 KiB
TeX
122 lines
4.2 KiB
TeX
\section{A/B/M/F drain services}
|
|
|
|
\label{sec:abmf}
|
|
%% ============================================================================
|
|
|
|
\subsection{Service set growth}
|
|
|
|
LP-132~\cite{lp132-quasarexec} (3.0 substrate) defined 12 GPU services.
|
|
QuasarSTM~4.0 grows the set to 16 by adding four new drains, each with
|
|
a dedicated device-resident ring and the same wave-tick budget contract
|
|
as the 3.0 services. The 16-service enum is the single normative source:
|
|
|
|
\begin{lstlisting}[language=C++]
|
|
enum class ServiceId : uint32_t {
|
|
// 3.0 substrate (12)
|
|
Ingress = 0,
|
|
Decode = 1,
|
|
Crypto = 2,
|
|
DagReady = 3,
|
|
Exec = 4,
|
|
Validate = 5,
|
|
Repair = 6,
|
|
Commit = 7,
|
|
StateRequest = 8,
|
|
StateResp = 9,
|
|
CertLane = 10,
|
|
CertOut = 11,
|
|
|
|
// 4.0 production (4 new) -- A/B/M/F
|
|
AdaptiveSchedule = 12,
|
|
BridgeAttest = 13,
|
|
MarketAuction = 14,
|
|
FiberCheckpoint = 15,
|
|
|
|
Count = 16,
|
|
};
|
|
\end{lstlisting}
|
|
|
|
\subsection{AdaptiveSchedule (A)}
|
|
|
|
\textbf{Inputs:} ConflictSpec records (\S\ref{sec:conflictspec}),
|
|
LaneClass classifications (\S\ref{sec:laneclass}), predictor entries.
|
|
\textbf{Outputs:} per-tx scheduling decision routed to
|
|
\texttt{Exec} (with declared lane sets attached) or to the speculative
|
|
fallback path.
|
|
|
|
The drain consults the source-priority order
|
|
(\S\ref{sec:conflictspec}) and emits the highest-priority
|
|
\texttt{ConflictSpec} per transaction. It also publishes the per-tx
|
|
predictor confidence so the validator can attribute later repair
|
|
events to predictor mis-prediction.
|
|
|
|
\subsection{BridgeAttest (B)}
|
|
|
|
\textbf{Inputs:} cross-chain attestation messages from the omnichain
|
|
bridge (LP-016 / LP-017). \textbf{Outputs:} per-message verifier
|
|
result onto the \texttt{Commit} ring.
|
|
|
|
BridgeAttest is the GPU-side termination of the cross-chain attestation
|
|
path: messages arrive as bundles of source-chain commitments and
|
|
threshold signatures, and the drain runs the same on-device BLS /
|
|
Corona / Groth16 verifiers (\S\ref{sec:verifiers}) against the
|
|
cross-chain VK published in the BridgeRoundDescriptor extension.
|
|
|
|
This is what makes 4.0 capable of executing bridge commits in the same
|
|
wave-tick budget as ordinary EVM, rather than spilling them to a
|
|
separate host-side path.
|
|
|
|
\subsection{MarketAuction (M)}
|
|
|
|
\textbf{Inputs:} order-append \texttt{DeferredOp}s
|
|
(\S\ref{sec:tier3}) and cancel ops within an auction batch window.
|
|
\textbf{Outputs:} \texttt{AuctionMatch} entries that reduce to the
|
|
batch-cleared book state at horizon time.
|
|
|
|
MarketAuction is the precompile-level batched matching path for hot
|
|
DEX lanes. It reduces an entire batch of order/cancel events into a
|
|
single match-and-clear transformation, eliminating the storm of
|
|
SSTORE conflicts that would otherwise hit the order book level lanes.
|
|
The output is a single \texttt{AuctionMatch} reducer entry per market
|
|
per round.
|
|
|
|
\subsection{FiberCheckpoint (F)}
|
|
|
|
\textbf{Inputs:} fiber state at \texttt{CALL}-frame return,
|
|
explicit checkpoint opcode, or cold-state suspend
|
|
(\S\ref{sec:checkpoints}). \textbf{Outputs:}
|
|
\texttt{FiberCheckpoint} records into the fiber's per-fiber checkpoint
|
|
ring.
|
|
|
|
This drain owns checkpoint emission and GC. It runs at
|
|
\texttt{CALL}-frame boundaries and at end-of-round to recycle older
|
|
checkpoints back into the per-round arena.
|
|
|
|
\subsection{Wave-tick budget}
|
|
|
|
Each new drain has a fixed per-tick budget enforced by the wave-tick
|
|
kernel. The budgets are tuned per backend; the cross-backend
|
|
determinism harness (\S\ref{sec:determinism}) verifies that any
|
|
budget setting that produces output also produces byte-identical
|
|
outputs across Metal and CUDA.
|
|
|
|
\begin{table}[ht]
|
|
\centering
|
|
\small
|
|
\caption{Per-tick budgets for the four new 4.0 services
|
|
(default; tunable).}
|
|
\label{tab:abmf-budgets}
|
|
\begin{tabular}{ll}
|
|
\toprule
|
|
\textbf{Service} & \textbf{Budget per tick} \\
|
|
\midrule
|
|
\texttt{AdaptiveSchedule} & 4096 ConflictSpec admissions \\
|
|
\texttt{BridgeAttest} & 256 attestation verifications \\
|
|
\texttt{MarketAuction} & 64 markets per auction tick \\
|
|
\texttt{FiberCheckpoint} & 4096 checkpoint emits + GC events \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\end{table}
|
|
|
|
%% ============================================================================
|