← all runs
Cohere-100M — 2026-09-02
r8i.16xlarge · peak recall 0.9965 @ 25 QPS
“Is HNSW build practical on 16xlarge?”
QPS at recall
| Recall | Meerkat | pgvector HNSW (halfvec) | Δ vs best competitor |
|---|
| 0.90 | 529 | 303 | +75% |
| 0.95 | 240 | 127 | +89% |
| 0.99 | 52 | — (max r=0.985) | — |
Latency at recall
| Meerkat | pgvector HNSW (halfvec) | Δ vs best competitor |
|---|
| @ recall 0.90 | | | |
| p50 | 1.88 ms | 3.33 ms | -44% |
| p95 | 2.29 ms | 4.48 ms | -49% |
| p99 | 2.57 ms | 5.41 ms | -52% |
| @ recall 0.95 | | | |
| p50 | 4.13 ms | 8.14 ms | -49% |
| p95 | 4.94 ms | 10.06 ms | -51% |
| p99 | 5.51 ms | 11.67 ms | -53% |
| @ recall 0.99 | | | |
| p50 | 19.31 ms | — | — |
| p95 | 22.54 ms | — | — |
| p99 | 24.70 ms | — | — |
All values are taken at exactly the stated recall: each configuration's own sweep is log-interpolated between its two surrounding measured points (measured values are used verbatim on an exact hit) and the best configuration at that recall is shown — values are never interpolated across different configurations. “—” means no single sweep spans that recall; nothing is extrapolated. Recall is distance-based (ann-benchmarks, tie-aware). Insert/update/delete workloads are not measured.
Interpolated on config — Meerkat @0.90: mkt-defaults · pgvector HNSW (halfvec) @0.90: hnsw-m24-halfvec · Meerkat @0.95: mkt-defaults · pgvector HNSW (halfvec) @0.95: hnsw-m24-halfvec · Meerkat @0.99: mkt-defaults
Build comparison
| Meerkat | pgvector HNSW (halfvec) | Δ vs best competitor |
|---|
| Build time | 2955 s | 151099 s | -98% |
| Index size | 24.9 GB | 179.2 GB | -86% |
Build stats are for the configuration backing each column's best operating point: Meerkat = mkt-defaults, prewarmed · pgvector HNSW (halfvec) = hnsw-m24-halfvec, prewarmed.
Plots
Pareto svg ·
Distance-recall svg ·
Per-config svg ·
Build times svg ·
Index sizes svgHost
| Instance type | r8i.16xlarge |
| Hostname | ip-172-31-76-48 |
| CPU | Intel(R) Xeon(R) 6975P-C (64 vCPU) |
| Memory | 496 GB |
| OS | Ubuntu 26.04.1 LTS |
| Kernel | 7.0.0-1011-aws |
Versions
| rekall | 0.1.0 (1834b41 / main) |
| PostgreSQL | PostgreSQL 18.6 (Ubuntu 18.6-1.pgdg26.04+2) on x86_64-pc-linux-gnu, compiled by gcc (Ubuntu 15.2.0-16ubuntu1) 15.2.0, 64-bit |
| meerkat | 0.1.0 (bdffa10) |
| pg_prewarm | 1.2 |
| vector | 0.8.6 |
Search
Index builds
| Index | Type | Build time | Size | Workers |
|---|
| hnsw-m24-halfvec | pgvector-hnsw | 151098.8s | 179.2 GB | 32 |
| mkt-defaults | meerkat | 2955.1s | 24.9 GB | 32 |
Index parameters
hnsw-m24-halfvec: {"ef_construction":200,"encoding":"halfvec","m":24}
mkt-defaults: {"boundary_epsilon":null,"centroid_compression":null,"centroid_fastscan":null,"fan_out":null,"fastscan":null,"fastscan_bits":null,"nlist":null,"rerank_vectors":null,"soar_lambda":null}
Instance settings (observed)
instance type r8i.16xlarge
vCPU (observed) 64
memory (observed) 495.8 GiB
cpu model Intel(R) Xeon(R) 6975P-C
kernel 7.0.0-1011-aws
data directory /var/lib/postgresql/18/main
mount point /
filesystem ext4
media unknown
device root
provisioned IOPS — not an EBS volume
PostgreSQL settings
checkpoint_completion_target = 0.9
cluster_name = 18/main
data_directory = /var/lib/postgresql/18/main
effective_cache_size = 384 GB
effective_io_concurrency = 256
huge_pages = try
jit = off
maintenance_io_concurrency = 256
maintenance_work_mem = 204 GB
max_connections = 100
max_parallel_maintenance_workers = 32
max_parallel_workers = 64
max_wal_size = 64 GB
random_page_cost = 1.1
server_version = 18.6 (Ubuntu 18.6-1.pgdg26.04+2)
shared_buffers = 128 GB
wal_level = replica
work_mem = 1 GB
