We had a report from a user on a machine with a 20 core CPU group that a hot lock in EventPipe (EventPipeBufferManager::m_lock) was causing excessive spinning in EventPipeBufferManager::AllocateBufferForThread. There was conflicting evidence about whether or not the system was sustaining high event load. Regardless of the specifics of that case, we need to validate that EventPipe scales well.

We should create a performance test that will EventPipe use from a large number of cores (>= 20) reading with both EventListener and IPC. We should then compare a low load case (1 event/sec/thread) with a high load case (maximum sustainable event rate on all threads) and compare the latency distribution in calls to WriteEvent. A first stab at a goal is probably that P50 latency increases no more than 10% and P99 latency increases no more than 50%.