Re: Improve eviction algorithm in ReorderBuffer

On Wed, 28 Feb 2024 at 11:40, Amit Kapila <amit.kapila16@gmail.com> wrote:
>
> On Mon, Feb 26, 2024 at 7:54 PM Masahiko Sawada <sawada.mshk@gmail.com> wrote:
> >
>
> A few comments on 0003:
> ===================
> 1.
> +/*
> + * Threshold of the total number of top-level and sub transactions
> that controls
> + * whether we switch the memory track state. While the MAINTAIN_HEAP state is
> + * effective when there are many transactions being decoded, in many systems
> + * there is generally no need to use it as long as all transactions
> being decoded
> + * are top-level transactions. Therefore, we use MaxConnections as
> the threshold
> + * so we can prevent switch to the state unless we use subtransactions.
> + */
> +#define REORDER_BUFFER_MEM_TRACK_THRESHOLD MaxConnections
>
> The comment seems to imply that MAINTAIN_HEAP is useful for large
> number of transactions but ReorderBufferLargestTXN() switches to this
> state even when there is one transaction. So, basically we use the
> binary_heap technique to get the largest even when we have one
> transaction but we don't maintain that heap unless we have
> REORDER_BUFFER_MEM_TRACK_THRESHOLD number of transactions are
> in-progress. This means there is some additional work when (build and
> reset heap each time when we pick largest xact) we have fewer
> transactions in the system but that may not be impacting us because of
> other costs involved like serializing all the changes. I think once we
> can try to stress test this by setting
> debug_logical_replication_streaming to 'immediate' to see if the new
> mechanism has any overhead.

I ran the test with a transaction having many inserts:

         | 5000     | 10000   |  20000  | 100000    |  1000000   | 10000000
------- |-----------|------------|------------|--------------|----------------|----------------
Head | 26.31 | 48.84 | 93.65  | 480.05  |  4808.29   | 47020.16
Patch |  26.35  | 50.8   | 97.99  | 484.8    |  4856.95   | 48108.89

The same test with debug_logical_replication_streaming= 'immediate'

         | 5000     | 10000   |  20000  | 100000    |  1000000   | 10000000
------- |-----------|------------|------------|--------------|----------------|----------------
Head | 59.29   |  115.84  |  227.21 | 1156.08   |  11367.42   |  113986.14
Patch | 62.45  |  120.48  |  240.56 | 1185.12   |  11855.37   |  119921.81

The execution time is in milliseconds. The column header indicates the
number of inserts in the transaction.
In this case I noticed that the test execution with patch was taking
slightly more time.

Regards,
Vignesh