Re: Adding basic NUMA awareness

On 9/11/25 10:32, Tomas Vondra wrote:
> ...
>
> 8) I've realized some of the TAP tests occasionally fail with
>
> ERROR: no unpinned buffers
>
> and I think I know why. Some of the tests set shared_buffers to a very
> low value - like 1MB or even 128kB, and StrategyGetBuffer() may search
> only a single partition (but not always). We may run out of unpinned
> buffers in that one partition.
>
> This apparently happens more easily on rpi5, due to the weird NUMA
> layout (there are 8 nodes with memory, but getcpu() reports node 0 for
> all cores).
>
> I suspect the correct fix is to ensure StrategyGetBuffer() scans all
> partitions, if there are no unpinned buffers in the current one. On
> realistic setups this shouldn't happen very often, I think.
>
> The other issue I just realized is that StrategyGetBuffer() recalculates
> the partition index over and over, which seems unnecessary (and possibly
> expensive, due to the modulo). And it also does too many loops, because
> it used NBuffers instead of the partition size. I'll fix those later.

Here's a version fixing this issue (in the 0006 part). It modifies
StrategyGetBuffer() to walk through all the partitions, in a round-robin
manner. The way it steps to the next partition is a bit ugly, but it
works and I'll think about some better way.

I haven't done anything about the other issue (the one with huge pages
reserved on NUMA nodes, and SIGBUS).

regards

--
Tomas Vondra