Currently each QueryKnobSnapshots point to a shared threadlocal clone in attempt to minimize the ref counter contention. This has two undesired effects:

• Increased memory usage scaled linearly to the thread count.
• Increased likelihood of being evicted from the LLC.

The proposed fix to address both issues is to use the shared_ptr aliasing constructor to separate the control and data blocks of QueryKnobSnapshots:

• The data pointer should point towards a shared global region
• The control pointer should point to a thread local instance.