Load Balancing

NUMA Load Balancing

Problem Description

Current Solaris scheduler chooses thread's home lgroup at thread creation time
based on load across different nodes. There is no mechanism to correct initial placment afterwords.

Prototype load balancing monitor

The following approach can be used to correct long-term load inbalance:

• Observe load for each lgroup over some period of time
• If any disbalance is found, find the best candidate for migration
• Rehome the thread 
• Migrate its memory as well

Here is the prototype Perl script that does that.