giraph-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ravikant Dindokar <ravikant.i...@gmail.com>
Subject Compute and network time for a superstep in Giraph
Date Sun, 29 May 2016 18:54:20 GMT
Hi Giraph user,

How can I compute, for each worker, for each superstep:

1. time spent in compute() method

2. time spent explicitly for sending messages (network cost)

and

synchronization time ( waiting for all worker to finish )

I used -Dgiraph.metrics.enable=true argument , but unable to interpret its
output:

output is something like this for a superstep:

--- METRICS: superstep 25 ---
*superstep time*
  mean: 236.875 ms
  slowest: 241 ms from orion-05.local_39
  fastest: 215 ms from orion-03.local_10
*compute all partitions*
  mean: 123.65 ms
  slowest: 157 ms from orion-03.local_25
  fastest: 101 ms from orion-01.local_18
*network communication time*
  mean: 0.0 ms
  slowest: 0 ms from orion-03.local_5
  fastest: 0 ms from orion-03.local_5
*time to first message*
  mean: 0.0 us
  slowest: 0 us from orion-03.local_5
  fastest: 0 us from orion-03.local_5
*wait requests time*
  mean: 727.25 us
  slowest: 4212 us from orion-03.local_15
  fastest: 164 us from orion-04.local_32

what is the meaning of each metric reported? is compute all partitions +
network communication time + synchronization overhead =superstep time or is
network communication time is included in the compute time itself?



Thanks
Ravikant

Mime
View raw message