This repository contains the reference eBPF/XDP drivers, Triton GPU lookup kernels, and microbenchmarking suites for the paper:
"Radix Economy and Balanced Ternary Microarchitectures: Resolving the Memory Wall in Line-Rate Post-Quantum Consensus and Nanoscale Computing"
Target Venues: SOSP / OSDI / ISCA / ASPLOS
64-Byte L1D Cache-Resident Frame : Compresses a 128-node consensus vote bitmask to 26 bytes ($3^5 = 243 \le 256$ ), enabling the entire synchronous descriptor (epoch + BLAKE3 accumulator + status flags) to fit within exactly one 64-byte L1D cache line. 2. Deterministic Wire Latency : Evaluated via eBPF XDP at 100GbE line-rate (99.2 Mpps aggregate across 8 queues, 12.4 Mpps/core) with$p50 = 40.0\text{ ns}$ and$p99.9 = 60.0\text{ ns}$ , safely below the$80.5\text{ ns}$ frame budget. 3. State-Crypt Separation : Decouples the 64-byte synchronous consensus frame from asynchronous ML-DSA-44 (NIST FIPS 204) signature verification offloaded over 2MB hugepage lock-free SPSC rings. 4. Conflict-Free GPU Decompression : Triton kernel maps 5-trit packed bytes into FP16 ternary weights with zero shared-memory bank conflicts using single-cycle hardware broadcast addressing.
xdp_ternary_filter.c- Production eBPF XDP C driver for line-rate packet parsing, SipHash-2-4 pre-authentication, monotonic epoch tracking, and fast-path quorum accumulation.triton_lut_kernel.py- Triton GPU kernel for high-throughput 5-trit decompression on Tensor Cores.benchmark_harness.py- Microarchitectural evaluation reproducing the latency and throughput ablations across 64-byte ternary and 96-byte binary frames.LICENSE- MIT License.
clang -O2 -target bpf -c xdp_ternary_filter.c -o xdp_ternary_filter.o
ip link set dev eth0 xdpgeneric obj xdp_ternary_filter.o sec xdp
python3 triton_lut_kernel.py
python3 benchmark_harness.py
@article{grimm2026radix,
title={Radix Economy and Balanced Ternary Microarchitectures: Resolving the Memory Wall in Line-Rate Post-Quantum Consensus and Nanoscale Computing},
author={Grimm, Justin},
year={2026}
}
MIT License - Copyright (c) 2026 Justin Grimm.