ECG HPCA Paper

ECG Successor HPCA Paper

This is the only current wiki page for the ECG successor architecture. The public paper name is intentionally still open; the implementation remains under the ECG_* namespace.

Important prior-publication boundary

The preliminary paper, ECG: Expressing Locality and Prefetching for Optimal Caching in Graph Structures, appeared in the archival IEEE IPDPSW 2024 proceedings (pp. 520–525, DOI 10.1109/IPDPSW59749.2024.00094).

HPCA does not permit a substantially similar archival-workshop submission. The new paper therefore requires:

1. materially distinct contributions;
2. citation and disclosure of the workshop paper;
3. a contribution-delta statement;
4. written guidance from the HPCA PC chairs before registration.

A new name improves differentiation but does not replace this requirement.

New architecture beyond the workshop paper

• K2 records: dest32 | epoch1_16 | epoch2_16.
• Adaptive replacement: epoch-first PR, degree-first BFS/SSSP, RRIP-first BC/CC.
• StreamShield placement: private-cache fills and LLC hits remain normal; bypassed LLC misses do not allocate.
• Request-bound ISA: ecg.load2 and ecg.stream.load2.
• Three-simulator artifact: cache_sim, gem5, and Sniper exact delivery and victim-decision gates.
• Full accounting: demand misses, total traffic, reserved P-OPT capacity, simulated time, and instruction count.

Current evidence

• Corrected real-graph cache_sim attribution: K2 77.3%, StreamShield 22.7%.
• gem5 synthetic mechanism cell: StreamShield improves fused K2 by 13.03%.
• Sniper synthetic mechanism cell: StreamShield improves fused K2 by 0.65% at identical instruction count.
• Overall detailed-simulator superiority over P-OPT is not claimed until the real-graph Sniper matrix completes.

Required paper comparison

Every reported table includes:

LRU  SRRIP  GRASP  charged P-OPT  K2  K2+StreamShield

The canonical policy labels are:

ECG:K2
ECG:K2_STREAMSHIELD

Canonical artifact

• Paper SSOT: research/ecg-hpca/
• Manifest: final_paper_manifest.json
• Runner: flows/paper_run.py
• Matrix engine: roi_matrix.py
• Slurm shards: slurm/make_slurm_shards.py

Canonical run

python3 scripts/experiments/ecg/flows/paper_run.py \
  --profile streamshield_sniper_realgraph \
  --run-dir results/ecg_experiments/final_paper_runs/ecg_successor_webgoogle \
  --no-build

Generated results/ data is not committed.