Select a host in the left side bar to open in Perfetto Select "ALL_HOSTS" to see the aggregated statistics table
Megascale Stats (all hosts)
This tool provides insights into the Megascale Stats.
(1) Rendezvous Name is the name assigned to the collective.
(2) Recv Op Name and Send Op Name is TPU recv-done op name and send op name.
(3) Slack Time is the network independent time the collective has to transmit the data.
(4) Observed Duration is the interval between the start of the send op to the end of the corresponding recv-done op.
(5) Stall Duration is the duration of time the collective spends in send/send-done/recv/recv-done ops.
(6) Occurrences is the total number of occurrences for each collective in the profiled duration.
(7) Aggregated Total Stall is the total stall duration of the collective in the profiled duration.
(8) Data Transmitted Size is the total data that needs to be transmitted over the network.
(9) Required Bandwidth is the bandwidth required to transmit the data in the provided slack.