nv.stats.latencymanager.samplesize

The global default size used for capturing latencies. Latency Latencies stats are collected in a ring buffer which is sampled by the stats thread at each collection interval.

This should be sized large enough such that datapoints aren't missed, but not so large that it adversely affects processor cache performance.

Default value: nv.stats.series.samplesize 

nv.stats.series.samplesize

If the number of datapoints collected in a stats interval exceeds this size the computation for histographical data will be lossy, increasing the value will reduce loss of datapoints, but results in greater overhead in stats collection in terms of both memory usage and pressure on the process caches.

Default value: 10240

nv.msg.latency.stats

This global property instructs the platform to collect latency statistics for messages passing through various points in the process pipeline.

when not enabled many latency stats will not be available in heartbeats.

nv.link.network.stampiots

Instructs low level socket I/O stamp input/output times on written data.

This should be enabled to capture store latencies based on wire time, or for certain latencies in the direct message bus binding.