WL#9387: InnoDB: Group purging of rows by table ID

Affects: Server-8.0   —   Status: Complete

Since the introduction of multi-threaded purge the purge threads work on random
undo log records. If different purge threads all purge from the same table this
leads to higher contention on the dict_index_t::lock (RW-LOCK).

To reduce dict_index_t::lock contention, we should group the undo rows by table
ID and then parcel the work to the purge threads based on the table id.

This is a feature request from Facebook Inc.
NF1 - Improvement in performance expected in 5.6 when purging rows from several
tables simultaneously. In 5.7+ the dict_index_t::lock improvements mitigate the
This is a very low level small internal change. It should improve performance
when lots of purge threads attempt to purge the rows from the same table.

1. Read the undo log records and group by table ID
2. Assign the undo log rows grouped during stage #1 to the purge threads
map{table_id, list {row}}

foreach row in undo_log
  parse table_id from row


# The aim is for each thread to work on a dedicated table
foreach list_of_rows in map
  foreach thread in threads
     thread.work += list_of_rows

Rest as usual.