- Infos im HLRS Wiki sind nicht rechtsverbindlich und ohne Gewähr -
- Information contained in the HLRS Wiki is not legally binding and HLRS is not responsible for any damages that might result from its use -

Workspace migration: Difference between revisions

From HLRS Platforms
Jump to navigationJump to search
Line 111: Line 111:
As a rule of thumb you get 1 GB/sec per node, but this is dependent on the number of files you are coping as well on the size of them.
As a rule of thumb you get 1 GB/sec per node, but this is dependent on the number of files you are coping as well on the size of them.


Again, pcp does NOT parallelize the copy operation itself, but the copy operations.  
Again, pcp does NOT parallelize a single copy operation itself, but the number of copy operations are distributed over the nodes.  


  > cat pcp.qsub
  > cat pcp.qsub

Revision as of 12:10, 8 December 2016

User migration on new workspaces

In December 2016 the workspaces installed in 2011 with the Cray Xe6 System Hermit will be replaced. In preparation for this task, users have to migrate their workspaces onto the replacement filesystems. Run the command ws_list -a on a frontend system to display the path for all your workspaces, if path names matches workspaces in following table, this workspace needs to be migrated.

File System alias mounted on
ws1 univ_1 /lustre/cray/ws1
ws2 ind_1 /lustre/cray/ws2
ws3 univ_2 /lustre/cray/ws3
ws4 res_1 /lustre/cray/ws4
ws5 ind_2 /lustre/cray/ws5
ws6 res_2 /lustre/cray/ws6

before you start

Migration for large amount of data consumes a lot of IO ressources. Please review and remove data which you do not need any more.

How to proceed ( Version 1)

  • Users have got access to the replacement Workspaces. To find out which one, try following command:
    • ws_allocate –F ws7 test_ws 5 # if this command run successful, you should prepare your Jobs using this workspace and submit all new compute Jobs utilizing this workspace.
    • If above command fails, following command should work: ws_allocate –F ws8 test _ws 5 # if not contact your project supervisor.
  • Run all new submitted Jobs within workspaces in the new location.
  • Migrate data from the “old” location into the fresh created workspace (please double check this target directory is located in either ws7 or ws8 directory tree).

to migrate we suggest following command: rsync -a Old_ws new_ws

  • On November 7th 2016 the default will be changed. Please ensure your jobs are using the new workspace directory
  • On December 7th the “old” workspaces ws1, … ws6 will be disconnected from the Cray system. The Filesystems will be available on the frontend systems for data migration until 11th January 2017
  • January 15th 2017 all data on the old filesystems will be deleted.


How to proceed ( Version 2)

  • On October, 25th 2016 new workspaces will be allocated on the replacment filesystems. Available workspace will be listed also on the old filesystems.
  • workspaces on old filesystems could not be extended.
  • if you have to migrate data from workspaces located in a different filesystem, do not use the mv command to transfer data. For large amount of data, this will fail. We recomment using following command: rsync -a ??? Old_ws new_ws
  • take care when you create new batch jobs. If you have to migrate your workspace from from an old filesystem to the new location, this take some time to transfer large amount of data. Do not run any Job while the migration process is active. This may result in incosistant data.
  • On December 7th the “old” workspaces ws1, … ws6 will be disconnected from the Cray system. The Filesystems will be available on the frontend systems for data migration until 11th January 2017
  • January 15th 2017 all data on the old filesystems will be deleted.


Using a parallel copy for data transfer

pcp is a python based parallel copy using MPI. It can only be run on compute nodes via aprun.

pcp is similar to cp -r; simply give it a source directory and destination and pcp will recursively copy the source directory to the destination inparallel.

pcp has a number of useful options; use pcp -h to see a description.

This program traverses a directory tree and copies files in the tree in parallel. It does not copy individual files in parallel. It should be invoked via aprun.

Basic arguments

If run with the -l flag or -lf flags pcp will be stripe aware.

-l will cause stripe information to be copied from the source files and directories.

-lf will cause all files and directories on the destination to be striped, regardless of the striping on the source.

Striping behavior can be further modified with -ls and -ld.

-ls will set a minimum file size. Files below this size will not be striped, regardless of the source striping.

-ld will cause all directories to be unstriped.

-b C: Copy files larger than C Mbytes in C Mbyte chunks

Algorithm

pcp runs in two phases:

Phase I is a parallel walk of the file tree, involving all MPI ranks in a peer-to-peer algorithm. The walk constructs the list of files to be copied and creates the destination directory hierarchy.

In phase II, the actual files are copied. Phase II uses a master-slave algorithm. R0 is the master and dispatches file copy instructions to the slaves (R1...Rn).

Job Script example

Here is an example of y job script.

You have to change the SOURCEDIR and TARGETDIR according to your setup. Also the number of nodes and wallclock time should be adjusted.

As a rule of thumb you get 1 GB/sec per node, but this is dependent on the number of files you are coping as well on the size of them.

Again, pcp does NOT parallelize a single copy operation itself, but the number of copy operations are distributed over the nodes.

> cat pcp.qsub
#!/bin/bash
#PBS -N IO_copy_test
#PBS -l nodes=8
#PBS -l walltime=0:30:00
#PBS -joe

cd $PBS_O_WORKDIR

module load tools/python/2.7.8

SOURCEDIR=<YOUR SOURCE DIRECTORY HERE>
TARGETDIR=<YOUR TARGET DIRECTORY HERE>

sleep 5 
nodes=$(qstat -f $PBS_JOBID | awk -F: '/Resource_List.nodes/ {print $1 }' | awk -F= '{print $2}')
let cores=nodes*24

/usr/bin/time -p aprun -n $cores -N24 -d1  pcp -l -ls 1048576  -b 4096 $SOURCEDIR $TARGETDIR
>

Output of a run with the script

R0: All workers have reported in.
Starting 192 processes.
Will copy lustre stripe information.
Files larger than 4096 Mbytes will be copied in parallel chunks.
Will not stripe files smaller than 1.00 Mbytes 

Starting phase I: Scanning and copying directory structure...
Phase I done: Scanned 115532 files, 1007 dirs in 00 hrs 00 mins 01 secs (106900 items/sec).
115532 files will be copied.

Starting phase II: Copying files...
Phase II done.

Copy Statisics:
Rank 1 copied 7.00 Gbytes in 839 files (38.17 Mbytes/s)
Rank 2 copied 6.37 Gbytes in 825 files (34.75 Mbytes/s)
...
Rank 190 copied 7.84 Gbytes in 495 files (42.78 Mbytes/s)
Rank 191 copied 7.25 Gbytes in 784 files (39.63 Mbytes/s)
Total data copied: 1.47 Tbytes in 115606 files (7.09 Gbytes/s)
Total Time for copy: 00 hrs 03 mins 31 secs
Warnings 0
Application 6324961 resources: utime ~4257s, stime ~4005s, Rss ~33000, inblocks ~3148863732, outblocks ~3148201243
real 259.90
user 0.02
sys 0.01

Operation of the workspaces will be changed:

  • To create a workspace or extend an existent workspace, an interactive shell is necessary.
  • Due to a drop of performance on high usage of quota, no job of any group member will be scheduled for computation as long as the group quota exceeds 80%. All blocked group members get a notice by E-mail (if a valid address is registered)
  • accounting
  • werden wir eine längere Laufzeit zulassen? Denke wir sollten, schon wegen 4 Wochen Urlaub... wie wäre es mit 60 Tagen???

Please read https://kb.hlrs.de/platforms/index.php/Storage_usage_policy