Tuning MySQL Cluster for NUMA ( eg after HW upgrade ) with ThreadConfig
(Doc ID 2527187.1)
Last updated on FEBRUARY 27, 2020
Applies to:MySQL Cluster - Version 7.2 and later
Information in this document applies to any platform.
This Document will explain the impact of Non-uniform_memory_access ( NUMA ) in a latency sensitive environment, specifically as MySQL Cluster.
It extends upon the previous KM ( which is a required reading component for this Document )
Optimising MySQL Cluster ndbmtd with ThreadConfig (Doc ID 2497622.1)
Production grade tuning is beyond scope for this DOC, due to complexity and unknown environments, however it can be used as an introduction to the topic.
This example will extend the example from (Doc ID 2497622.1) ;
The Server has recently been upgraded from its previous configuration of 1 CPU and 64G RAM, to 2 CPUs and an additional 64G.
CPU is SMT capable ( Simultaneous_multithreading ) with 8 cores ( 16 threads).
This could be considered as a typical enterprise config for a Server : now having 128G RAM , with both CPU sockets having 4x16Gs installed "locally" ( meaning lower latency than accessing the *other* memory bank ).
The above is explicitly mentioned due to the impact that logical cores / SMT and the distance ( between memory bank and CPU socket ) can have on a latency sensitive system, if the access is "non local".
Assuming the pre requisite reading has been covered, as outlined in Optimising MySQL Cluster ndbmtd with ThreadConfig (Doc ID 2497622.1) this example will cover
- performance expectations
- deciding how to enable ndbmtd to make use of the additional resources
- further discussion of ThreadConfig options
The scenario in one where a cluster has been performing nominally using an existing manual ThreadConfig variable for its ndbmtd processes;
Physical core / Logical core - Thread assignment ( 1 Socket )
0,8 - Reserved for OS
1,9 - LDM0,LDM4
2,10 - LDM1,LDM5
3,11 - LDM2,LDM6
4,12 - TC0,
5,13 - TC1,
6,14 - SEND0,RECV0
7,15 - MAIN,IO0
Due to proactive capacity planning and an expectation that the dataset will increase along with the number of users, the cluster ndbmtd nodes have been upgraded as stated.
The goal now is to maximize the benefit received from the new upgrades.
To view full details, sign in with your My Oracle Support account.
Don't have a My Oracle Support account? Click to get started!
In this Document
|First boot expectations ( after hardware upgrade validated )|
|Recommendation for OS NUMA configuration changes|
|Effective use of CPU/RAM resources , post hardware upgrade.|
|Further discussion of MySQL Cluster performance options|