make_global_df: Extract global data (once-per-rank data) from an...
In marcpaterno/timestampr: Analysis of MPI timestamp logs

Description Usage Arguments Details Value

Each row in the returned dataframe corresponds to one rank in the program.

1	make_global_df(dx)

`dx`	an eventselection raw dataframe

The dataframe columns are:

rank: the MPI rank on which the call was made
start: time at which the rank started running
post_dataset: time at which the HEPnOS dataset creation finished
post_block_configurations: time at which creation of EventSelectionConfigs finished
post_broadcast: time at which broadcast of EventSelectionConfigs is finished
pre_decompose: time before making DIY blocks
post_decompose: time after making DIY blocks
pre_execute_block: time before execution of blocks
post_execute_block: time after execution of blocks
pre_create_partners: time before creation of DIY merge partners
pre_reduction: time before DIY reduction starts
post_reduction: time after DIY reduction finishes
finish: time after output is written (see note 3 below)
makeds: duration for creation of HEPnOS dataset
calcbc: duration for making EventSelectionConfigs (see note 1 below)
broadcast: duration of broadcast (see note 1 below)
prep: duration of prep work before making blocks
makeblocks: duration of block creation
makelambda: duration creation of prefetcher and of the lambda that executes the blocks
executeblock: duration of the execution of all blocks
makepartners: duration of the creation of merge partners
reduction: duration of all reduction rounds
output: duration of writing result file (see note 2 below)
total: total running duration

Note 1: Only the local group leaders calculate block configurations, so for most ranks the 'calcbc' value will be close to zero. For these ranks, the 'broadcast' time will include the time spent waiting for the group leaders to do their calculations of block configurations.

Note 2: Only rank 0 writes the output file; others ranks should have close to zero time for 'output'.

Note 3: The timingdata files are written *after* the finish timestamp is recorded.

a tibble containing global data

marcpaterno/timestampr documentation built on Jan. 12, 2022, 6:41 a.m.