Sustained heavy swapping can grind execution to a halt, but today users get no direct warning from the node when this happens, and this caveat is not documented.
We can’t easily detect heavy swap pressure in a reliable, cross-platform way, but we can document what swapping is and why it can severely degrade IBD performance.
Note: An earlier version of this PR attempted to detect and warn on heavy swapping (Linux-only), but it was changed to documentation based on review feedback.
The following sections might be updated with supplementary metadata relevant to reviewers and maintainers.
See the guideline for information on the review process. A summary of reviews will appear here.
No conflicts as of last run.
Instead of warning how about just flushing the dbcache?
We could do both, thanks for the idea. Anything else?
Anything else?
dbcache is the main memory hog. We could also reduce mempool space on the fly and have bitcoind evict low paying txs. Not sure that’s feasible in the current code though. Dropping peers would be next I think.
but they all depend on the user realizing that there’s a problem.
I don’t think a one-time log in debug.log is going to solve that problem though (especially not during any period of large logging activity)?
I think as a first step flush the cache. That would probably fix the issue in almost all cases.
In the reproducer above, that’s not the case, but in reality I would expect that, yes.
Currently, this is only triggered a single time, so that likely would not be enough to “fix the issue in almost all cases”. And flushing every time there’s swapping pressure could cause more problems if they end up triggering or worsening each other somehow.
Probably want to flush before we get into swap territory, so that’s independent from this PR IMO.
But also, using swap does not mean the system is struggling. Linux, at least, can move inactive pages into swap just to improve I/O caching, and I wouldn’t be surprised if reindexing triggered this behaviour.
What I had in mind is to have the simplest heuristic to help users who are complaining of IBD taking months.
That seems like something much better solved by documentation (“minimum node requirements”). Swapping isn’t, in general, something we can react to – it can be normal/desirable behaviour (some other process on the system has loaded 1GB of information into memory that it isn’t accessing frequently, so 700MB of it swapped out, eg).
Document what swapping is and why it can severely degrade `bitcoind` performance.