Section: 4.4 Using the scanalyze

"Receive lines" has the following fields:

<Comm><rank> recv from <from>(<worldFrom>):<commonFields>

where

 

<Comm>

is the communicator being used

<rank>

is the rank within <Comm>

<from>

is the rank within <Comm>

<worldFrom>

is the rank within MPI_COMM_WORLD

"Send-lines" has the following fields:

<Comm><rank> send to <to>(<worldTo>):<commonFields> where

<Comm>

is the communicator being used

<rank>

is the rank within <Comm>

<to>

is the rank within <Comm>

<worldTo>

is the rank within MPI_COMM_WORLD

The <commonFields> are as follows:

!<count>!<avrLen>!<zroLen>!<inline>!<eager>!<transporter>! where

<count>

is the number of sends/receives

<avrLen>

is the average length of messages in bytes

<zroLen>

is the number of messages sent/received using the zero-bytes

 

mechanism

<inline>

is the number of messages sent/received using the inline

 

mechanism

<eager>

is the number of messages sent/received using the eagerbuffer

 

mechanism

<transporter>

is the number of messages sent/received using the transporter

 

mechanism

More details on the different mechanisms can be found in “Description of Scali MPI Connect” on page 11.

4.4 Using the scanalyze

Tracing and timing the image processing example above produced little data, and interpreting the data posed little problem. However, most applications run for longer time, with correspondingly larger logs as a result; output from tracing and timing with Scali MPI Connect easily amounts to megabytes of data.

In order to extract information from the huge amount of data, Scali has developed a simple analysis tool called scanalyze. This analysis tool accept output from SMC applications run with certain predefined trace and timing variables set.

4.4.1 Analysing all2all

The all2all program in /opt/scali/examples/bin is a simple communication benchmark, but tracing and timing it produces massive log files. For example running

user% SCAMPI_TRACE=”-f arg;timing” mpimon ./all2all -- r1 r2

on a particular system produced a 2354159 byte log file, while running

Scali MPI Connect Release 4.4 Users Guide

43

Page 55
Image 55
Escali 4.4 Using the scanalyze, Commrank recv from fromworldFromcommonFields, Commrank send to toworldTocommonFields where