Improve memory use

[matrixbot]

This issue was originally created by @ptman at matrix-org/dendrite#1805.

Description:

Dendrite can use a lot of RAM at times and this can result in it getting killed by the oom-killer.

Keywords: ram, memory, oom, resource use, efficiency, killed

[matrixbot]

This comment was originally posted by @ptman at matrix-org/dendrite#1805 (comment).

#1804 could improve this a lot

[matrixbot]

This comment was originally posted by @ptman at matrix-org/dendrite#1805 (comment).

Caching helps a lot in synapse matrix-org/dendrite#1102

[matrixbot]

This comment was originally posted by @kegsay at matrix-org/dendrite#1805 (comment).

Caching bloats baseline memory, it doesn't help it. It helps reduce CPU cycles.

There are several problems which can cause Dendrite to gobble up memory:

• complicated state resolution. State resolution looks at historical state as well as the current state so older rooms or rooms with lots of membership churn will reduce performance because all those events need to be extracted from the database.
• resolving state with lots of prev_events. If a room has a lot of servers and they constantly fork the room, when the forks get merged we need to do state resolution. Even small rooms (with lots of forks) can reduce performance due to needing to extract the events multiple times. We can probably improve this with an event cache so we can have pointers to the same structure.

An "easy" fix to all this is to not federate, as all these issues only appear once you start doing state resolution which only happens when servers disagree or need to work out the room state.

[matrixbot]

This comment was originally posted by @ptman at matrix-org/dendrite#1805 (comment).

Yes, caching uses more memory. But counter-intuitively it does sometimes help with synapse, as the cached result uses less memory than the process of re-calculating the value.

[matrixbot]

This comment was originally posted by @neilalexander at matrix-org/dendrite#1805 (comment).

It would be interesting to get some feedback on how much recent changes have helped here. Version 0.4.0 specifically contains some pretty huge improvements on memory usage, especially when dealing with incoming federation events, calculating memberships and running state resolution, so hopefully it is noticeable.

[matrixbot]

This comment was originally posted by @neilalexander at matrix-org/dendrite#1805 (comment).

@ptman I'd be curious to know if you think this is still a problem, especially given recent optimisations in v0.6.x. There are also additional things that have been merged onto main today that should help even more (#2226, matrix-org/gomatrixserverlib#292). I'd like to close this if it's no longer a noticeable issue.

[matrixbot]

This comment was originally posted by @ptman at matrix-org/dendrite#1805 (comment).

Sorry, I lost my dendrite installation. I'll have to deploy a new one at some point when I have a chance. I'll close this for now.