When gathering cpu stats collector locks cache for each cpu(core) update. With high number of logical cpus (500+) it needs 500+ cache locks per second just to gather the data. This should be optimized by gathering statistics for all cpus locally and then flushing that data with single cache lock.