Scalable Coherence: Existing Techniques Landscape

Adapted from Table I — techniques that improve at most 2 of {Storage, Latency, Traffic} at the expense of others

📦 Bloom Filter Dir

[67][70] Probabilistic tracking. Saves storage but false-positive invalidations + longer lookup/eviction.

✓ Storage✗ Latency— Traffic

📐 Coarse-Grain Track

[8][41] Region-level tracking. Saves traffic when lines in same region share similarly. Extra region hw.

— Storage✗ Latency✓ Traffic

📏 Multi-Grain Coh.

[12][14][66] Coh. tracked at granularities coarser than cache line. Saves storage+traffic, adds latency.

✓ Storage✗ Latency✓ Traffic

🗂️ Cuckoo Directory

[16] Cuckoo-hash-based organization. Smaller size for same entries, but insertion latency increases.

✓ Storage✗ Latency— Traffic

💻 OS-Assisted

[13][15][28][31][61] HW within nodes, OS page-level across nodes. Saves storage, OS invocations cost.

✓ Storage✗ Latency✗ Traffic

📤 Overflow into LLC

[10][29][57] WayPoint, ZeroDEV, TinyDir. Dir entries overflow into LLC. Saves storage, less LLC capacity.

✓ Storage✗ Latency— Traffic

🏠 Home Delegation

[11][50][51] Direct node-to-node, bypass home indirection. Reduces latency+traffic, needs metadata.

— Storage✓ Latency✓ Traffic

📡 Network Optim.

[26] VTC — hw-assisted multicast via virtual trees in routers. Saves latency+traffic, extra router hw.

— Storage✓ Latency✓ Traffic

🔮 TX Prediction

[24][46] PATCH, Push Multicast — predict TX patterns. Reduces latency, hw cost for recording.

— Storage✓ Latency✓ Traffic

📋 Reduced Inval ACKs

[33] ACKwise — records N sharers, expects N ACKs after broadcast. Saves storage, no silent eviction.

✓ Storage— Latency✗ Traffic

🔄 DRAM Replication

[45] Dvé — duplicates dir+memory for fault tolerance. Less latency (closer copy), doubles storage.

✗ Storage✓ Latency✗ Traffic

⭐ Dorado (This Work)

TLH + Dynamic Apport. + SetOverflow — simultaneously addresses all three dimensions.

✓ Storage✓ Latency✓ Traffic

Every prior technique improves at most 2/3 dimensions. Dorado is the first to address all three simultaneously.