Skip to content

Scalability Curves

This page documents HEBBS’s performance characteristics at different memory scales. All measurements use the hebbs-bench tool on standardized hardware.

ParameterValue
HardwareAWS m6i.2xlarge (8 vCPU, 32 GB RAM)
Storagegp3 EBS, 3000 IOPS
HNSW M16
HNSW ef_construction200
HNSW ef_search100
EmbeddingBGE-small-en-v1.5, CPU
Concurrent clients100
Memory Countp50p95p99
100K0.8 ms1.5 ms2.1 ms
500K1.5 ms3.2 ms4.5 ms
1M2.3 ms4.8 ms6.2 ms
5M3.8 ms7.1 ms8.9 ms
10M4.5 ms8.3 ms9.8 ms
50M6.2 ms12.1 ms15.3 ms
100M8.1 ms16.4 ms19.7 ms

Similarity recall scales logarithmically — doubling the memory count adds roughly 1ms to p99.

Memory Countp50p95p99
100K0.3 ms0.6 ms0.9 ms
1M0.5 ms1.1 ms1.5 ms
10M0.8 ms1.8 ms2.4 ms
100M1.2 ms2.5 ms3.3 ms

Temporal recall remains fast at all scales due to B-tree logarithmic lookup.

Memory CountAvg Depthp50p95p99
100K3.21.1 ms2.5 ms3.8 ms
1M3.51.4 ms3.1 ms5.2 ms
10M3.81.8 ms4.2 ms7.1 ms

Causal recall depends more on graph depth than total memory count.

Memory CountSingle-threadedBatched (32)
100K620/sec3,400/sec
1M580/sec3,200/sec
10M520/sec2,900/sec
100M450/sec2,500/sec

Write throughput degrades gradually as the HNSW index grows (insertion becomes slightly more expensive).

Memory CountDisk UsageRAM (RSS)HNSW Index Size
100K180 MB420 MB95 MB
500K850 MB1.8 GB470 MB
1M1.7 GB3.5 GB940 MB
5M8.3 GB16 GB4.7 GB
10M16.5 GB31 GB9.4 GB
50M82 GB150 GB47 GB
100M165 GB290 GB94 GB

The HNSW index is the dominant memory consumer, accounting for ~55-60% of total RAM usage.

  • Similarity recall stays under 10ms p99 up to 10M memories on a single node.
  • Temporal and causal recall remain fast at all tested scales.
  • Write throughput degrades gradually but stays above 2,500 batched writes/sec even at 100M.
  • Memory usage is dominated by the HNSW index. Plan ~3 GB RAM per million memories.
  • Disk usage is roughly ~1.6 GB per million memories with default compression.