Support global memory control

[wshwsh12]

Feature Request

Is your feature request related to a problem? Please describe:

Currently, TiDB has a query-level memory control strategy mem-quota-query, which triggers Cancel when the memory usage of a single SQL exceeds mem-quota-query. However, there is currently no global memory control strategy. When TiDB has a certain degree of inaccuracy due to multiple SQLs whose memory usage does not exceed mem-quota-query or memory tracking, it will lead to high memory usage or even OOM. Therefore, we need a means that can be observed from the outside to determine whether the memory usage of the current system is normal, and when there is a problem, try to control its memory not to continue to grow, so as to reduce the risk of process crashes.

Design Docs

Todo List

• Kill the SQL with the most memory usage: When the memory usage of the TiDB instance reaches tidb_server_memory_limit, kill the SQL with the largest memory usage in the current system. *: support global memory control for tidb #37794
  • Soft limit to trigger spill disk, hard limit to tirgger cancel. Change the behavior of mem-quota-query #37104
  • Support bytesize format and percentage format for system variable. sessionctx: support bytes format and percentage format for tidb_server_memory_limit #38655
  • Deprecate serverMemoryQuota and inherit the values from old cluster. session: tidb_server_memory_limit inherits the value of server-memory-quota when upgrading #38628
• Take the initiative to trigger GC: When the memory usage of the TiDB instance reaches tidb_server_memory_limit * tidb_server_memory_limit_gc_trigger, try to trigger GC *: Support to trigger GC when memory usage is large. #38179
• Introduce some memory tables: Introduce some memory tables to observe the memory usage of the current system. executor: add some memory tables to observe memory usage #38452

Test

• Function test https://github.com/pingcap/endless/pull/1120
• Performance test
• Compatibility test