Hints via cudaMemAdvise

Prefetch moves data now; cudaMemAdvise instead tells the driver how you plan to use a region so it can place pages smartly over time.

You pass the pointer, byte count, an advice flag, and a device id. The driver folds that hint into its future migration decisions.

Use cudaMemAdviseSetReadMostly for data many devices read but rarely write. The driver can keep read-only copies on each side to avoid migration.

With cudaMemAdviseSetPreferredLocation you nominate a home device for the pages. They resist drifting away even when another side touches them.

cudaMemAdviseSetAccessedBy says a device will reach this region, so the driver maps it there. Access goes over the bus instead of moving pages.

When two devices fight over the same pages, the right advice ends the bounce. You stabilize placement so data stops migrating back and forth.

Unlike a one-shot prefetch, advice persists on the region until you change it. It shapes every migration choice for the life of the buffer.

Every Set flag has a matching Unset form to revoke the hint. Clear advice when an access pattern changes so it does not mislead the driver.

The two tools pair well. Set PreferredLocation to define the home, then prefetch to fill it, and migrations settle into a steady pattern.

Remember the name: advice is a suggestion, not a guarantee. The driver may override it, so always profile to confirm the placement helped.

You now hold three levers: one managed pointer, prefetch for timing, and advise for placement. Together they make Unified Memory both easy and fast.

Let us confirm how cudaMemAdvise differs from prefetching.

You learned cudaMemAdvise hints like ReadMostly, PreferredLocation, and AccessedBy to guide page placement. Pair it with prefetch and profile. Well done! 🎯