Space-Analog AI Inference Characterization
Research project studying LLM inference and KV cache behavior under space-like operating conditions.
How do LLM inference metrics — throughput, latency, KV cache hit rate, output correctness — degrade when compute operates under space-analog constraints: radiation-elevated BER, passive thermal management only, and intermittent network links?
$5,000 research grant — TCP networking and space-analog compute study.
Characterize TCP behavior under space-analog link conditions using Python-based link emulation. Uses tc/netem for link degradation modeling.
Modular compute unit with passive cooling, radiation-component selection, thermal interface design. Physical platform for Track 1 experiments.
src/
emulation/ # Space-analog link and hardware condition emulation
inference/ # LLM inference benchmarking (vLLM, MLC-LLM)
kvcache/ # KV cache behavior under fault conditions
networking/ # TCP characterization under degraded links
thermal/ # Thermal simulation and throttling models
experiments/
tcp_characterization/ # TCP failure modes under space-analog links
inference_bench/ # Inference throughput/latency under faults
kvcache_bench/ # KV cache hit rate, eviction, corruption
thermal_sim/ # Passive cooling thermal dynamics
results/
baseline/ # Standard datacenter conditions
space_analog/ # Space-like conditions
degraded/ # Intermediate degradation levels
| Paper | Relevance |
|---|---|
| PagedAttention (Kwon et al., 2023) | KV cache paged memory — radiation fault isolation |
| DCTCP (Alizadeh et al., 2010) | TCP baseline for datacenter links |
| FlashAttention (Dao et al., 2022) | Attention tiling on constrained memory hierarchy |
| AWQ (Lin et al., 2024) | Quantization under bandwidth-constrained hardware |
| ShadowKV (Sun et al., 2024) | KV cache offload under memory constraints |
| Message Ferries (Zhao et al., 2004) | Intermittent connectivity networking |
| Snoop Protocol (Balakrishnan et al., 1995) | Link-layer TCP shielding for unreliable links |
| L4S (Briscoe et al., 2023) | Scalable congestion signaling for low-latency links |
- TCP throughput degradation curve vs BER (0 to 10^-3)
- Inference throughput (tokens/sec) vs thermal throttle level
- KV cache hit rate vs DRAM bit error rate
- Output faithfulness (ROUGE/perplexity) vs radiation level
- Latency P50/P95/P99 across all conditions
- SIGCOMM, NSDI (networking track)
- MLSys, OSDI (inference systems track)
- Hot Topics in Space Networking (if applicable)