Priority refill for the buffer pool free list¶

In highly-concurrent I/O-bound workloads the following situation may happen:

1. Buffer pool free lists are used faster than they are refilled by the LRU cleaner thread.
2. Buffer pool free lists become empty and more and more query and utility (i.e. purge) threads stall, checking whether a buffer pool free list has became non-empty, sleeping, performing single-page LRU flushes.
3. The number of buffer pool free list mutex waiters increases.
4. When the LRU manager thread (or a single page LRU flush by a query thread) finally produces a free page, it is starved from putting it on the buffer pool free list as it must acquire the buffer pool free list mutex too. However, being one thread in up to hundreds, the chances of a prompt acquisition are low.

This is addressed by delegating all the LRU flushes to the to the LRU manager thread, never attempting to evict a page or perform a LRU single page flush by a query thread, and introducing a backoff algorithm to reduce buffer pool free list mutex pressure on empty buffer pool free lists. This is controlled through a new system variable innodb_empty_free_list_algorithm.

variable innodb_empty_free_list_algorithm¶

Command Line:	Yes
Config File:	Yes
Scope:	Global
Dynamic:	Yes
Values:	legacy, backoff
Default Value:	backoff

When legacy option is set, server will used the upstream algorithm and when the backoff is selected, Percona implementation will be used.

Multi-threaded LRU flusher¶

Percona Server for MySQL features a true multi-threaded LRU flushing. In this scheme, each buffer pool instance has its own dedicated LRU manager thread that is tasked with performing LRU flushes and evictions to refill the free list of that buffer pool instance. Existing multi-threaded flusher no longer does any LRU flushing and is tasked with flush list flushing only.

• All threads still synchronize on each coordinator thread iteration. If a particular flushing job is stuck on one of the worker threads, the rest will idle until the stuck one completes.
• The coordinator thread heuristics focus on flush list adaptive flushing without considering the state of free lists, which might be in need of urgent refill for a subset of buffer pool instances on a loaded server.
• LRU flushing is serialized with flush list flushing for each buffer pool instance, introducing the risk that the right flushing mode will not happen for a particular instance because it is being flushed in the other mode.

The following InnoDB metrics are no longer accounted, as their semantics do not make sense under the current LRU flushing design: buffer_LRU_batch_flush_avg_time_slot, buffer_LRU_batch_flush_avg_pass, buffer_LRU_batch_flush_avg_time_thread, buffer_LRU_batch_flush_avg_time_est.

The need for InnoDB recovery thread writer threads is also removed, consequently all associated code is deleted.

Parallel doublewrite buffer¶

The legacy doublewrite buffer is shared between all the buffer pool instances and all the flusher threads. It collects all the page write requests into a single buffer, and, when the buffer fills, writes it out to disk twice, blocking any new write requests until the writes complete. This becomes a bottleneck with increased flusher parallelism, limiting the effect of extra cleaner threads. In addition, single page flushes, if they are performed, are subject to above and also contend on the doublewrite mutex.

To address these issues Percona Server for MySQL uses private doublewrite buffers for each buffer pool instance, for each batch flushing mode (LRU or flush list). For example, with four buffer pool instances, there will be eight doublewrite shards. Only one flusher thread can access any shard at a time, and each shard is added to and flushed completely independently from the rest. This does away with the mutex and the event wait does not block other threads from proceeding anymore, it only waits for the asynchronous I/O to complete. The only inter-thread synchronization is between the flusher thread and I/O completion threads.

The new doublewrite buffer is contained in a new file, where all the shards are contained, at different offsets. This file is created on startup, and removed on a clean shutdown. If it’s found on a crashed instance startup, its contents are read and any torn pages are restored. If it’s found on a clean instance startup, the server startup is aborted with an error message.

The location of the doublewrite file is governed by a new innodb_parallel_doublewrite_path global, read-only system variable. It defaults to xb_doublewrite in the data directory. The variable accepts both absolute and relative paths. In the latter case they are treated as relative to the data directory. The doublewrite file is not a tablespace from InnoDB internals point of view.

The legacy InnoDB doublewrite buffer in the system tablespace continues to address doublewrite needs of single page flushes, and they are free to use the whole of that buffer (128 pages by default) instead of the last eight pages as currently used. Note that single page flushes will not happen in Percona Server for MySQL unless innodb_empty_free_list_algorithm is set to legacy value.

The existing system tablespace is not touched in any way for this feature implementation, ensuring that cleanly-shutdown instances may be freely moved between different server flavors.

Version Specific Information¶

• 8.0.12-1 Feature ported from Percona Server for MySQL 5.7

Other Reading¶

• Page cleaner thread tuning
• Bug #74637 - make dirty page flushing more adaptive
• Bug #67808 - in innodb engine, double write and multi-buffer pool instance reduce concurrency
• Bug #69232 - buf_dblwr->mutex can be splited into two