DataPower · An open benchmark

DataPowerUseful work per unit of hardware

5 systems. Same data. Same operations. 2 CPU cores, 8 GB of memory.*

0100200300400500600DataPower300K1M3MData elements, log scale

* Measurements from Nov. 1, 2026. See the specs.

  • 520DP

    DataPower at 3M elements

    TETRAbase scored 520. PostgreSQL 16.15 scored 198 and SQLite 3.53.4 scored 112.*

  • 6.5×

    Score growth, 300K to 3M elements

    TETRAbase rose from 80 to 520 as the data grew tenfold. PostgreSQL 16.15 rose from 98 to 198.*

  • 376ops/s

    Useful throughput at 3M

    TETRAbase completed 376 correct operations a second, adjusted for slow responses, with 4 clients. PostgreSQL 16.15 completed 230.*

  • 2.9ms

    CPU per operation at 3M

    TETRAbase used 2.9 ms of processor time per operation. PostgreSQL 16.15 used 7.5 ms and SQLite 3.53.4 used 9.1 ms.*

* Measurements from Nov. 1, 2026. See the specs.

01 · The results

Every result, side by side.

Choose a dataset size to compare all five databases. Click a database's name to follow it through the charts below.

Size
Score
DataPower*520.2197.5112.367.225.5
DP-Read, reads only1,000.0140.3178.367.736.4
DP-Write, writes only193.0682.9176.01,254.25.3
Work
Data elements, E3 million3 million3 million3 million3 million
Useful throughput, T375.8 ops/s230.1 ops/s113.1 ops/s54.9 ops/s27.7 ops/s
Clients at best throughput44441
Latency score at best throughput, Q0.650.890.560.730.65
Hardware
Resource units, R2.1672.0562.0152.2282.414
CPU cores22222
Average memory, M1.33 GiB0.41 GiB0.11 GiB1.67 GiB3.31 GiB
Disk, D0.23 GiB0.91 GiB0.19 GiB3.92 GiB0.11 GiB
Disk per data element82 bytes325 bytes66 bytes1,402 bytes39 bytes
CPU time per operation2.94 ms7.54 ms9.09 ms25.89 ms27.89 ms
Minimum memory8 GiB1 GiB8 GiB2 GiB4 GiB
Setup
Setup points0141027
Setup multiplier, O1.001.701.501.101.35
Checks
Wrong answers00000
Writes that did not read back00000
Sources
Test runReport →results.jsonReport →results.jsonReport →results.jsonReport →results.jsonReport →results.json
Date
1 November 2026
Server
AMD EPYC 7502P with 32 cores and 252 GiB of memory, running Linux 6.12 and Podman 4.3.1
Per database
2 CPU cores and 8 GiB of memory, in a container
Data
DP-Core, a generated dataset (seed 42), at 300K, 1M and 3M data elements
Workload
80% everyday operations, 15% queries on columns with no index, 5% worst-case inputs
Test profile
Quick: 1, 4, 16 and 64 clients for 5 seconds each, then 2 minutes alternating busy and idle
Runs
One per database and size
Rules
DataPower 2026.1, measured with dpbench
  • * These results fall short of the full DataPower standard in four ways. The quick profile shortens every timed step. The standard takes the median of three runs, and we ran each test once. The standard sizes are 1M, 10M, 100M and 1B elements, so 300K and 3M are extra. The containers ran with extra privileges, because the server's kernel lacks a feature they need. Quick runs vary by about a third from one run to the next, so trust large gaps more than small ones.
  • Every run passed the checks: no database returned a wrong answer or lost a write.
  • We tested a development build of TETRAbase, which the run reports call TETRA v0. The other four databases are public releases.

02 · Test runs

Each test run has its own report.

Every report gives the score and its parts, response times for each kind of operation, memory and disk use, and a timeline from the cold start to the final checks. The results.json under each score holds the raw data.

System300K elements1M elements3M elements
TETRAbase v0development build80.2DPReport → results.json264.6DPReport → results.json520.2DPReport → results.json
PostgreSQL 16.15relational97.7DPReport → results.json172.0DPReport → results.json197.5DPReport → results.json
SQLite 3.53.4embedded, relational73.9DPReport → results.json109.2DPReport → results.json112.3DPReport → results.json
Neo4j 5.26Community, graph33.1DPReport → results.json54.7DPReport → results.json67.2DPReport → results.json
lance-csr 0.5.1embedded, graph on Lance6.3DPReport → results.json12.3DPReport → results.json25.5DPReport → results.json

Compare all fifteen runs

The comparison charts all five databases from 300K to 3M elements, from 1 to 64 clients, and from a cold start.

Open the comparison →

03 · How the score works

DataPower divides useful work by the hardware and setup behind it.

TETRA CA PBC maintains DataPower and its test harness, dpbench. A database scores higher when it holds more data and completes more useful work each second, and lower when it needs more hardware or more setup.

DP= E × TR × O
  • E

    Data elements

    E counts the data, in millions of elements. Each table row is one element, and so is each link from one row to another.

    E = elements ÷ 1,000,000

  • T

    Useful throughput

    T counts correct operations per second, discounted for slow responses. We run 1, 4, 16 and 64 clients and measure X, correct answers per second, and Q, a latency score from 0 to 1. T is the highest X × Q.

    T = max (X × Q)

  • R

    Resource units

    R counts the hardware. Each CPU core adds 1, each 8 GiB of memory adds 1, and each 200 GiB of disk adds 1.

    R = cores + M ÷ 8 + D ÷ 200

  • O

    Setup

    O counts the setup. Each index, tuned setting, hand-sized cache, maintenance job, extra service or runtime adds 1 or 2 points, and each point adds 5%.

    O = 1 + 0.05 × points

At 3M data elements. Every database holds the same data, so E is equal, and the 2 CPU cores make up most of R. T and O decide the ranking.

  • Latency score, Q

    Each kind of operation has a time limit that 99% of its requests should meet. An operation within its limit scores 1. An operation over its limit scores the limit divided by its time, so an operation at twice the limit scores 0.5. Q multiplies these scores, weighted by how often each operation runs, so one slow operation pulls down the whole score.

  • Average memory, M

    M is the database's average memory. The harness samples it ten times a second while the database alternates between busy and idle minutes. A database that releases memory when idle scores better.

  • Disk, D

    D is the size of the data on disk, after loading or after the writes, whichever is larger.

  • DP-Read and DP-Write

    DP-Read and DP-Write apply the same formula to reads alone and to writes alone. DataPower covers the full mix.

Finding T

0200400600141664clientsoperations per secondT = 375.8X, correctanswers per secondX × Q, usefulthroughput

TETRAbase v0 at 3M did its most useful work with 4 clients: T = 375.8 operations a second. With 64 clients it answered 552.0 a second, but its latency score fell to 0.18.

The calculation for TETRAbase v0 at 3M data elements

Report →results.json
E3million elements
×
T375.8ops/s at 4 clients
÷
R2.167resource units
×
O1.000 setup points
=
DP520.2DataPower

T = X 581.6 × Q 0.646 = 375.8, with 4 clients, the peak in the chart above

R = 2 cores + 1.33 GiB ÷ 8 + 0.23 GiB ÷ 200 = 2.167

O = 1 + 0.05 × 0 points = 1.00

DP = 3 × 375.8 ÷ (2.167 × 1.00) = 520.2

Each test run, in seven steps

  1. Load. The harness loads the data into a container with 2 CPU cores and 8 GiB of memory.
  2. Cold start. The harness stops the database, clears the server's file cache, restarts the database and times its first queries.
  3. Warm-up. The harness runs the workload until throughput settles.
  4. Clients. The harness runs 1, 4, 16 and 64 clients and measures throughput and response times at each. This step gives T.
  5. Busy and idle. The harness alternates busy and idle minutes and samples memory. This step gives M.
  6. Minimum memory. The harness lowers the memory limit until performance drops, and reports that limit beside the score.
  7. Checks. The harness reads back every write and compares the database's settings with a fresh install. Setup that nobody declared still costs points.

The rules

  • Every database gets the same data, the same operations and the same hardware limits.
  • The harness checks every answer against a known correct answer. It rejects a run if more than 0.1% of any kind of operation returns a wrong answer, or if a single write fails to read back. A rejected run gets no score.
  • The harness runs every database from a fixed image, pinned by its digest, and takes every figure from the run's own records.
  • Each database runs with the settings its adapter declares, and every declared setting counts toward O.

Score your own database.

dpbench works with any database that can load the data and run the operations. Book a call, and we'll run it on yours.