Hi, thanks for looking into this.
On Sat, Oct 18, 2025 at 4:59 PM Kirill Reshke <reshkekirill@gmail.com> wrote:
>
> On Sat, 18 Oct 2025 at 12:50, Xuneng Zhou <xunengzhou@gmail.com> wrote:
> >
> > Hi Hackers,
>
> Hi!
>
> > The SnapBuildPurgeOlderTxn function previously used a suboptimal
> > method to remove old XIDs from the committed.xip array. It allocated a
> > temporary workspace array, copied the surviving elements into it, and
> > then copied them back, incurring unnecessary memory allocation and
> > multiple data copies.
> >
> > This patch refactors the logic to use a standard two-pointer, in-place
> > compaction algorithm. The new approach filters the array in a single
> > pass with no extra memory allocation, improving both CPU and memory
> > efficiency.
> >
> > No behavioral changes are expected. This resolves a TODO comment
> > expecting a more efficient algorithm.
> >
>
> Indeed, these changes look correct.
> I wonder why b89e151054a0 did this place this way, hope we do not miss
> anything here.
I think this small refactor does not introduce behavioral changes or
breaks given constraints.
> Can we construct a microbenchmark here which will show some benefit?
>
I prepared a simple microbenchmark to evaluate the impact of the
algorithm replacement. The attached results summarize the findings.
An end-to-end benchmark was not included, as this function is unlikely
to be a performance hotspot in typical decoding workloads—the array
being cleaned is expected to be relatively small under normal
operating conditions. However, its impact could become more noticeable
in scenarios with long-running transactions and a large number of
catalog-modifying DML or DDL operations.
Hardware:
AMD EPYC™ Genoa 9454P 48-core 4th generation
DDR5 ECC reg
NVMe SSD Datacenter Edition (Gen 4)
Best,
Xuneng
