About automatic memory management

AlloyDB Omni uses adaptive algorithms for memory management.

You can decide the upper limit of the shared buffer on starting AlloyDB Omni. If you don't set the upper limit, AlloyDB Omni automatically sets the shared buffer backing up size to 80% of system memory. The initial backing up size of the shared buffer can be different from the upper limit.

AlloyDB Omni consists of an intelligent memory worker which constantly monitors the memory status, and tunes shared buffer backing up size for best performance when caching data.

By default the shared_buffers parameter is set to 0 which is a special value that sets the upper limit of the size of the shared buffers cache to 80% of system memory. AlloyDB Omni starts at 10% of the shared_buffers upper limit. If shared_buffers is overridden by a custom value, then AlloyDB Omni respects the value as the upper limit of the shared_buffers size, and starts with that specified custom size.

Note: Memory worker can still decide to shrink, and later grow shared buffer backing up size at runtime.

To specify a custom size, edit the postgresql.conf configuration file. For example, you can set shared_buffers to 1GB using either of the following:

• docker run --name CONTAINER_NAME -e INITDB_ARGS="-c shared_buffers=1GB" $image

• docker run --name CONTAINER_NAME $image -c shared_buffers=1GB

  Replace CONTAINER_NAME with the name that you assigned to the AlloyDB Omni container when you installed it.

Optimize query performance

The default value of the shared_buffers parameter works for common scenarios.

However, you can tune the value for best performance. If you choose to rely on the default value of shared_buffers to deduce the shared buffer upper limit, then use the cgroup memory.max value to influence the computation.

Columnar engine memory

The dynamic shared_buffers is independent of the columnar engine memory. When the columnar engine is enabled, the dynamic shared_buffers size can be derived by subtracting the amount of memory used by the columnar engine from 80% of the total memory available to the system or cgroup.

Huge pages

Huge pages improve database performance. AlloyDB Omni manages huge pages explicitly if possible, otherwise it relies on the transparent huge pages (THP) feature of the operating system. If neither huge page type is supported, AlloyDB Omni falls back to 4k page, and prints a warning in docker container logs docker logs $container_name with specific instructions to set up huge pages. See Start AlloyDB Omni for instructions how to start a container.

The warning looks similar to the following:

HINT:  Please either execute the all-in-one setup script:
          docker run --rm --privileged $image setup-host
        OR manually execute:
          echo within_size | sudo tee /sys/kernel/mm/transparent_hugepage/shmem_enabled
          sudo sysctl -w vm.nr_overcommit_hugepages=1048576

Automatic memory management at runtime

AlloyDB Omni constantly monitors the system load, and adjusts its memory consumptions for better performance. Specifically, you might observe the following:

Dynamic shared_buffers size change

AlloyDB Omni increases the dynamic shared_buffers size when the system memory consumption is low, and decreases the size when the system memory consumption is high. To monitor the dynamic shared_buffers size, use:

CREATE EXTENSION IF NOT EXISTS g_memory;
SELECT g_dynamic_shared_size();

Termination of a PostgreSQL connection when the system is extremely low on memory

When AlloyDB Omni detects that the system is extremely low on memory, it attempts to delete the most memory-consuming PostgreSQL connections until the load falls back to a reasonable level. When such an event happens, AlloyDB Omni logs the following in docker container logs:

WARNING: Sending SIGTERM to pid=xxx NSpid=xxx (VA size = xxxMB) (RSS size = xxxMB)