darshan

Darshan has been turned on by default on Hopper for user as of Nov. 15, 2012

Description

Darshan is a light weight IO profiling tool capable of profiling POSIX IO, MPIIO and HDF5 IO. We encourage all users to turn on Darshan for their application running on Hopper. Darshan will not only help the users to identify IO bottleneck and improve performance, but also help NERSC to better understand the IO usage of its users and shape its future plans. 

Enable Darshan on Hopper

The Darshan module is loaded by default for all Hopper users. Simply recompile your code to allow Darshan to collect statistics. 

Darshan Results

Locating Results

All Darshan results are located in /scratch2/scratchdirs/darshanlogs/year/month/day/user-executablename_idPBS_JOBID_xxx.gz. You can use darshan-parser to see the details of the results. Note that each aprun command will produce a separate log file. 

Understanding Results

You can use the darshan-parser tool to analyze your log. To show how much data was read or written in a run:

darshan-parser --total logfile | grep BYTES_READ
darshan-parser --total logfile | grep BYTES_WRITTEN

To show the number of Read/Write operations in the run:

darshan-parser --total logfile | grep POSIX| grep READS
darshan-parser --total logfile | grep POSIX| grep WRITES 

You can also get a distribution of the transaction size, for POSIX read/write:

% darshan-parser --total logfile | grep SIZE_READ|grep -v AGG 
total_CP_SIZE_READ_0_100: 44
total_CP_SIZE_READ_100_1K: 12
total_CP_SIZE_READ_1K_10K: 16
total_CP_SIZE_READ_10K_100K: 12
total_CP_SIZE_READ_100K_1M: 28
total_CP_SIZE_READ_1M_4M: 8
total_CP_SIZE_READ_4M_10M: 8
total_CP_SIZE_READ_10M_100M: 12
total_CP_SIZE_READ_100M_1G: 12
total_CP_SIZE_READ_1G_PLUS: 4 

Or for MPIIO:

% darshan-parser --total logfile | grep SIZE_READ|grep AGG 
total_CP_SIZE_READ_AGG_0_100: 4
total_CP_SIZE_READ_AGG_100_1K: 12
total_CP_SIZE_READ_AGG_1K_10K: 16
total_CP_SIZE_READ_AGG_10K_100K: 12
total_CP_SIZE_READ_AGG_100K_1M: 20
total_CP_SIZE_READ_AGG_1M_4M: 8
total_CP_SIZE_READ_AGG_4M_10M: 8
total_CP_SIZE_READ_AGG_10M_100M: 12
total_CP_SIZE_READ_AGG_100M_1G: 12
total_CP_SIZE_READ_AGG_1G_PLUS: 4

Example

The following script will setup a test directory in $SCRATCH, automatically compile and run it, and point you to the profiling output. 

#if you use bash or sh
/usr/common/usg/darshan/example/rundarshanExample.sh

#if you use tcsh or csh
/usr/common/usg/darshan/example/rundarshanExample.csh

#if you don't know which shell you have
env|grep SHELL 

Overhead of Darshan

The overhead of Darshan is negligible to most of jobs. The plot below shows the overhead of darshan with respect to the job concurrency. The blue markers shows the overhead when each MPI task is reading/writing its own file. The red markers shows the overhead when all MPI tasks are read/writing the same file. For big jobs, it is suggested to use shared file IO (E.g. MPIIO) instead of file pre process. "Overhead" is defined as the time that Darshan uses to communicate the result and write the log file. The overhead of darshan should be un-noticeable for jobs with less than 10000 MPI tasks.  

Availability

Contact

If you have any question about Darshan at NERSC, please email consult@nersc.gov