heap / memory usage guidance vs number of vectors & dimensions

[tae898]

Hi,

Is there any documented or recommended guidance on JVM heap sizing or
memory usage as a function of:

• number of vectors (N)
• vector dimensionality (D)
• index parameters (e.g. maxDegree / beamWidth)
• compressed vs uncompressed vectors (PQ / BQ)

I couldn’t find this in the README or existing issues.

Even a rough order-of-magnitude estimate (e.g. approximate bytes per
vector for graph structures vs vectors, or typical heap requirements
observed in practice) would be very helpful for capacity planning and
deployment.

Is this documented somewhere already, or would it make sense to add
such guidance to the documentation?

Thanks.

[tae898]

Hi JVector team,

The ArcadeDB team ran experiments here: ArcadeData/arcadedb#3140

These tests aren’t a pure JVector benchmark—JVector runs inside ArcadeDB with a thin Python binding—but the logs and memory profile indicate most of the compute and RSS are on the JVector side. Overall, it’s very usable.

One concern is heap requirements during index build. If the heap is reduced below the thresholds in the report, build time balloons (to the point of being nearly unusable). Any guidance on reducing heap needs or keeping build times reasonable at lower heaps would be appreciated.

[tae898]

btw 4.0.0-rc.6 was used for this experiment.