Re: Direct I/O

On 4/19/23 10:11, Robert Haas wrote:
> On Tue, Apr 18, 2023 at 3:35 PM Greg Stark <stark(at)mit(dot)edu> wrote:
>> Well.... I'm more optimistic... That may not always be impossible.
>> We've already added the ability to add more shared memory after
>> startup. We could implement the ability to add or remove shared buffer
>> segments after startup. And it wouldn't be crazy to imagine a kernel
>> interface that lets us judge whether the kernel memory pressure makes
>> it reasonable for us to take more shared buffers or makes it necessary
>> to release shared memory to the kernel.
>
> On this point specifically, one fairly large problem that we have
> currently is that our buffer replacement algorithm is terrible. In
> workloads I've examined, either almost all buffers end up with a usage
> count of 5 or almost all buffers end up with a usage count of 0 or 1.
> Either way, we lose all or nearly all information about which buffers
> are actually hot, and we are not especially unlikely to evict some
> extremely hot buffer.

That has been my experience as well, although admittedly I have not
looked in quite a while.

> I'm not saying that it isn't possible to fix this. I bet it is, and I
> hope someone does.

I keep looking at this blog post about Transparent Memory Offloading and
thinking that we could learn from it:

https://engineering.fb.com/2022/06/20/data-infrastructure/transparent-memory-offloading-more-memory-at-a-fraction-of-the-cost-and-power/

Unfortunately, it is very Linux specific and requires a really up to
date OS -- cgroup v2, kernel >= 5.19

> I'm just making the point that even if we knew the amount of kernel
> memory pressure and even if we also had the ability to add and remove
> shared_buffers at will, it probably wouldn't help much as things
> stand today, because we're not in a good position to judge how large
> the cache would need to be in order to be useful, or what we ought to
> be storing in it.

The tactic TMO uses is basically to tune the available memory to get a
target memory pressure. That seems like it could work.

--
Joe Conway
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com