[Ideas] 关于提升Cloudberry高并发查询锁优化的建议

[saga1015]

Description

在给客户做POC测试的时候，刚好服务器CPU是华为鲲鹏的，在搜索如何提升查询性能的时候，看到在华为云的论坛里有这么一篇文章《GreenPlum高并发查询锁优化》链接是https://bbs.huaweicloud.com/blogs/317803。
其中提到了在鲲鹏服务器上s_lock的在高并发查询场景下影响性能的问题分析，以及提交的补丁。只不过补丁在这个文章里下载不到了。通过搜索在EDB的论坛里也有讨论这个问题的，还给出了补丁。
附件为补丁内容。我想是不是这个补丁也能合并到cbdb中，这样如果客户的环境是鲲鹏的，那性能就能得到提升。
cas.patch

Use case/motivation

No response

Related issues

No response

Are you willing to submit a PR?

• Yes I am willing to submit a PR!

[tuhaihe]

Hi, thanks for the feedback. Could you please share the link to the original source?

Any code contributed to Apache Cloudberry needs to have a clear origin and comply with the Apache licensing requirements to avoid any infringement risks.

By the way, did this actually work during your POC process, or just a idea? Thanks.

[saga1015]

https://postgrespro.ru/list/thread-id/2519320

[gfphoenix78]

Interesting! The patch adds a new implementation of spin lock, i.e. __sync_lock_test_and_set vs __atomic_compare_exchange_n.
__sync_lock_test_and_set performs better than __atomic_compare_exchange_n with low concurrency. But __atomic_compare_exchange_n behaves much better than __sync_lock_test_and_set when concurrency grows.

[gfphoenix78]

Tests 3 spinlock implementations on x86_64 and arm64:

1. c11 atomic flag (flag_.test_and_set(std::memory_order_acquire))
2. __atomic_compare_exchange_n
3. __sync_lock_test_and_set

The atomic implementations performs worse than the third implementation on arm64.
But on a small virtual host, the atomic implementations performs better. It's not clear
how it affects the performance in practice.