mirror of
https://github.com/lux-apps/papers.git
synced 2026-11-03 05:07:46 +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.
64 lines
2.6 KiB
TeX
64 lines
2.6 KiB
TeX
\section{ZAP: Zero-Copy Wire Format for Consensus Messages}
|
|
|
|
\label{sec:zap}
|
|
|
|
FPC's message complexity of $O(k \log n)$ makes per-message overhead a first-order concern. The default networking stack (gRPC + Protobuf) imposes substantial deserialization costs that dominate round-trip latency at high throughput. We replace Protobuf with the Zero-copy Application Protocol (ZAP), a binary wire format that enables direct memory access without deserialization.
|
|
|
|
\subsection{Design}
|
|
|
|
ZAP encodes consensus messages as fixed-layout binary structures with explicit field offsets. The receiver reads fields directly from the network buffer via offset arithmetic, eliminating the allocate-deserialize-copy pipeline of Protobuf.
|
|
|
|
\begin{lstlisting}[language=Go, caption=ZAP message layout for FPC vote]
|
|
// ZAP FPC Vote Message (72 bytes fixed)
|
|
// Offset 0: [32]byte TxID
|
|
// Offset 32: [4]byte Round (uint32, big-endian)
|
|
// Offset 36: [1]byte Opinion (0x00 = false, 0x01 = true)
|
|
// Offset 37: [1]byte Confidence (uint8)
|
|
// Offset 38: [2]byte Reserved
|
|
// Offset 40: [32]byte Signature (Ed25519 or BLS)
|
|
|
|
func (m *ZAPVote) TxID() ids.ID {
|
|
return ids.ID(m.buf[0:32]) // zero-copy: no allocation
|
|
}
|
|
|
|
func (m *ZAPVote) Round() uint32 {
|
|
return binary.BigEndian.Uint32(m.buf[32:36])
|
|
}
|
|
|
|
func (m *ZAPVote) Opinion() bool {
|
|
return m.buf[36] == 0x01
|
|
}
|
|
\end{lstlisting}
|
|
|
|
\subsection{Benchmark Results}
|
|
|
|
Benchmarks conducted on Apple M2 Ultra (24-core, 192GB) comparing ZAP against Protobuf for FPC consensus messages:
|
|
|
|
\begin{table}[h]
|
|
\centering
|
|
\begin{tabular}{lcccc}
|
|
\toprule
|
|
\textbf{Metric} & \textbf{Protobuf} & \textbf{ZAP} & \textbf{Improvement} \\
|
|
\midrule
|
|
Deserialization (ns/op) & 487 & 34 & $14.3\times$ \\
|
|
Serialization (ns/op) & 312 & 28 & $11.1\times$ \\
|
|
Allocations (allocs/op) & 7 & 0 & $\infty$ \\
|
|
Throughput (msg/s, single core) & 690K & 9.69M & $14.0\times$ \\
|
|
Message size (bytes, FPC vote) & 94 & 72 & $1.3\times$ smaller \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{ZAP vs.\ Protobuf for FPC consensus messages}
|
|
\label{tab:zap}
|
|
\end{table}
|
|
|
|
\subsection{Throughput Impact}
|
|
|
|
At $k = 20$ samples per round and 5 rounds to finality, each consensus decision requires $k \times 5 \times 2 = 200$ messages (query + response). With ZAP's 9.69M messages/second throughput on a single core:
|
|
|
|
\begin{equation}
|
|
\text{TPS}_{\text{theoretical}} = \frac{9.69 \times 10^6}{200} = 48{,}450 \text{ tx/s (single core)}
|
|
\end{equation}
|
|
|
|
This matches the empirical benchmarks in Table~\ref{tab:benchmarks} (45,000--52,000 TPS), confirming that ZAP removes serialization as the bottleneck and consensus message processing becomes network-bound rather than CPU-bound.
|
|
|