Revision as of 11:41, 11 July 2013

Introduction

The mpiblast module must be manually loaded before submitting any mpiblast jobs. Two examples below demonstrate how to setup and submit jobs to the mpi queue.

Version Selection

module unload openmpi intel; module load intel/11.0.083 openmpi/intel/1.4.2 mpiblast/1.6.0

Sample Problems

DROSOPH

Copy sample problem files (fasta database and input) from the /opt/sharcnet examples directory a directory under work as shown here. The fasta database used in this example can be obtained as a guest from NCBI here http://www.ncbi.nlm.nih.gov/guide/all/#downloads_ then clicking "FTP: FASTA BLAST Databases".

mkdir -p /work/$USER/samples/mpiblast/test1; rm /work/$USER/samples/mpiblast/test1/*
cd /work/$USER/samples/mpiblast/test1
cp /opt/sharcnet/mpiblast/1.6.0/examples/drosoph.in drosoph.in
gunzip -c /opt/sharcnet/mpiblast/1.6.0/examples/drosoph.nt.gz > drosoph.nt

Create hidden configuration file to define a Shared storage location between nodes and a Local storage directory available on each compute node where $USER should be replaced with your username as shown here:

[username@orc-login1:/work/roberpj/samples/mpiblast/test1] vi .ncbirc
[BLAST]
BLASTDB=/scratch/YourUserName/mpiblasttest1
BLASTMAT=/work/YourUserName/samples/mpiblast/test1
[mpiBLAST]
Shared=/scratch/YourUserName/mpiblasttest1
Local=/tmp

Format the database into 16 fragments under the following local scratch directory location:

mkdir /scratch/$USER/mpiblasttest1; rm -f /scratch/$USER/mpiblasttest1/*
cd /work/$USER/samples/mpiblast/test1
mpiformatdb -N 8 -i drosoph.nt -o T -p F -n /scratch/$USER/mpiblasttest1

For example ...

[roberpj@hnd20:/work/roberpj/samples/mpiblast/test1] mpiformatdb -N 8 -i drosoph.nt -o T -p F -n /scratch/roberpj/mpiblasttest1
Reading input file
Done, read 1534943 lines
Breaking drosoph.nt into 8 fragments
Executing: formatdb -p F -i drosoph.nt -N 8 -n /scratch/roberpj/mpiblasttest1/drosoph.nt -o T 
Created 8 fragments.
<<< Please make sure the formatted database fragments are placed in /scratch/roberpj/mpiblasttest1/ before executing mpiblast. >>>

[roberpj@hnd20:/scratch/roberpj/mpiblasttest1] ls
drosoph.nt.000.nhr  drosoph.nt.002.nin  drosoph.nt.004.nnd  drosoph.nt.006.nni
drosoph.nt.000.nin  drosoph.nt.002.nnd  drosoph.nt.004.nni  drosoph.nt.006.nsd
drosoph.nt.000.nnd  drosoph.nt.002.nni  drosoph.nt.004.nsd  drosoph.nt.006.nsi
drosoph.nt.000.nni  drosoph.nt.002.nsd  drosoph.nt.004.nsi  drosoph.nt.006.nsq
drosoph.nt.000.nsd  drosoph.nt.002.nsi  drosoph.nt.004.nsq  drosoph.nt.007.nhr
drosoph.nt.000.nsi  drosoph.nt.002.nsq  drosoph.nt.005.nhr  drosoph.nt.007.nin
drosoph.nt.000.nsq  drosoph.nt.003.nhr  drosoph.nt.005.nin  drosoph.nt.007.nnd
drosoph.nt.001.nhr  drosoph.nt.003.nin  drosoph.nt.005.nnd  drosoph.nt.007.nni
drosoph.nt.001.nin  drosoph.nt.003.nnd  drosoph.nt.005.nni  drosoph.nt.007.nsd
drosoph.nt.001.nnd  drosoph.nt.003.nni  drosoph.nt.005.nsd  drosoph.nt.007.nsi
drosoph.nt.001.nni  drosoph.nt.003.nsd  drosoph.nt.005.nsi  drosoph.nt.007.nsq
drosoph.nt.001.nsd  drosoph.nt.003.nsi  drosoph.nt.005.nsq  drosoph.nt.dbs
drosoph.nt.001.nsi  drosoph.nt.003.nsq  drosoph.nt.006.nhr  drosoph.nt.mbf
drosoph.nt.001.nsq  drosoph.nt.004.nhr  drosoph.nt.006.nin  drosoph.nt.nal
drosoph.nt.002.nhr  drosoph.nt.004.nin  drosoph.nt.006.nnd

Submit a short job with a 15m time limit on 8 plus 2 cores. If all goes well output results will be written to drosoph.out and the execution time will appear in ofile%J where %J is the job number:

cd /work/$USER/samples/mpiblast/test1
sqsub -r 15m -n 10 -q mpi --mpp=1G -o ofile%J mpiblast -d drosoph.nt -i drosoph.in
    -p blastn -o drosoph.out --use-parallel-write --use-virtual-frags

For example ...

[roberpj@hnd20:/work/roberpj/samples/mpiblast/test1] sqsub -r 15m -n 10 -q mpi --mpp=1G -o ofile%J mpiblast -d drosoph.nt -i drosoph.in -p blastn -o drosoph.out --use-parallel-write --use-virtual-frags
submitted as jobid 6966896

[roberpj@hnd20:/work/roberpj/samples/mpiblast/test1] cat ofile6966896.hnd50 
Total Execution Time: 1.80031

When submitting a mpiblast job on a cluster such as goblin that doesnt have an inifiniband interconnect better performance (at least double speedup) will be achieved running the mpi job on one compute node. For regular users of non-contributed hardware typically specify "-n 8" to reflect the max number of cores on a single node:

sqsub -r 15m -n 8 -N 1 -q mpi --mpp=4G -o ofile%J mpiblast -d drosoph.nt -i drosoph.in
           -p blastn -o drosoph.out --use-parallel-write --use-virtual-frags

Sample output results computed previously with BLASTN 2.2.15 [Oct-15-2006] are included in /opt/sharcnet/mpiblast/1.6.0/examples/ROSOPH.out to compare your newly generated drosoph.out file with.

2) UNIGENE

The main purpose of this example is to illustrate some additional options and switchs that maybe useful for debugging and for dealing with larger databases as described in official detail at http://www.mpiblast.org/Docs/Guide. The fasta database used in this example can also be downloaded from http://www.ncbi.nlm.nih.gov/guide/all/#downloads_ as a guest by clicking "FTP: UniGene" then entering the "Homo_sapiens" sub-directory. More information about UniGene alignments can be found at https://cgwb.nci.nih.gov/cgi-bin/hgTrackUi?hgsid=95443&c=chr1&g=uniGene_3 . As with Example1 above, for convenience all required files can simply be copied from the /opt/sharcnet examples subdirectory to work as shown here:

mkdir /work/$USER/samples/mpiblast/test2; rm /work/$USER/samples/mpiblast/test2/*
cd /work/$USER/samples/mpiblast/test2
cp /opt/sharcnet/mpiblast/1.6.0/examples/il2ra.in il2ra.in
gunzip -c /opt/sharcnet/mpiblast/1.6.0/examples/Hs.seq.uniq.gz > Hs.seq.uniq

Create hidden configuration file using the vi editor to define a Shared storage location between nodes and a Local storage directory available on each compute node as follows, where the Data directory is not yet populated or used in this example and hence can be omitted, where $USER should be replaced with your username as shown here:. If its desired the Local and Shared directories are the same then replace --copy-via=mpi with --copy-via=none as will be demonstrated in the below sqsub commands.

[username@orc-login1:/work/roberpj/samples/mpiblast/test2] vi .ncbirc
[NCBI]
Data=/opt/sharcnet/mpiblast/1.6.0/data
[BLAST]
BLASTDB=/work/YourUserName/mpiblasttest2
BLASTMAT=/work/YourUserName/samples/mpiblast/test2
[mpiBLAST]
Shared=/work/YourUserName/mpiblasttest2
Local=/tmp

Partition the database into 16 fragments under the following work directory location:

mkdir -p /scratch/$USER/mpiblasttest2; rm -f /scratch/$USER/mpiblasttest2/*
cd /work/$USER/samples/mpiblast/test1
mpiformatdb -N 16 -i Hs.seq.uniq -o T -p F                                          

Submit a couple of short jobs 15m time limit. If all goes well output results will be written to biobrewA.out and biobrewB.out and the execution time appear in corresponding ofile%J's where %J is the job number as per usual:

A) In this job submission fragment files are first copied from work to local /tmp before being used (appropriate if work is slow). Usage of the profile option is also shown in this example:

cd /work/$USER/samples/mpiblast/test2; rm -f oTime*
sqsub -r 15m -n 18 -q mpi -o ofile%J mpiblast  --use-parallel-write --copy-via=mpi -d Hs.seq.uniq -i il2ra.in -p blastn -o biobrew.out --time-profile=oTime

B) In this job submission fragment files are used inplace on work. Usage of the debug option is also shown in this example.

cd /work/$USER/samples/mpiblast/test2; rm -f oLog*
sqsub -r 15m -n 18 -q mpi -o ofile%J mpiblast --use-parallel-write --copy-via=none -d Hs.seq.uniq -i il2ra.in -p blastn -o biobrew.out --debug=oLog

Finally compare /opt/sharcnet/mpiblast/1.6.0/examples/BIOBREW.out computed previously with BLASTN 2.2.15 [Oct-15-2006] with your newly generated biobrew.out output file to verify the results and submit a ticket if there are any problems!

SUPPORTED PROGRAMS IN MPIBLAST

As described in http://www.mpiblast.org/Docs/FAQ mpiblast supports the standard blast programs http://www.ncbi.nlm.nih.gov/BLAST/blast_program.shtml which are reproduced here for reference:

blastp:     Compares an amino acid query sequence against a protein sequence database.
blastn:     Compares a nucleotide query sequence against a nucleotide sequence database.
blastx:     Compares a nucleotide query sequence translated in all reading frames against
            a protein sequence database.
tblastn:    Compares a protein query sequence against a nucleotide sequence database
            dynamically translated in all reading frames.
tblastx:    Compares the six-frame translations of a nucleotide query sequence against
            the six-frame translations of a nucleotide sequence database.

MPIBLAST BINARIES OPTIONS

[roberpj@orc-login1:/opt/sharcnet/mpiblast/1.6.0/bin] ./mpiblast -help
mpiBLAST requires the following options: -d [database] -i [query file] -p [blast program name]

[roberpj@orc-login1:/opt/sharcnet/mpiblast/1.6.0/bin] ./mpiformatdb --help
Executing: formatdb - 

formatdb 2.2.20   arguments:

  -t  Title for database file [String]  Optional
  -i  Input file(s) for formatting [File In]  Optional
  -l  Logfile name: [File Out]  Optional
    default = formatdb.log
  -p  Type of file
         T - protein   
         F - nucleotide [T/F]  Optional
    default = T
  -o  Parse options
         T - True: Parse SeqId and create indexes.
         F - False: Do not parse SeqId. Do not create indexes.
 [T/F]  Optional
    default = F
  -a  Input file is database in ASN.1 format (otherwise FASTA is expected)
         T - True, 
         F - False.
 [T/F]  Optional
    default = F
  -b  ASN.1 database in binary mode
         T - binary, 
         F - text mode.
 [T/F]  Optional
    default = F
  -e  Input is a Seq-entry [T/F]  Optional
    default = F
  -n  Base name for BLAST files [String]  Optional
  -v  Database volume size in millions of letters [Integer]  Optional
    default = 4000
  -s  Create indexes limited only to accessions - sparse [T/F]  Optional
    default = F
  -V  Verbose: check for non-unique string ids in the database [T/F]  Optional
    default = F
  -L  Create an alias file with this name
        use the gifile arg (below) if set to calculate db size
        use the BLAST db specified with -i (above) [File Out]  Optional
  -F  Gifile (file containing list of gi's) [File In]  Optional
  -B  Binary Gifile produced from the Gifile specified above [File Out]  Optional
  -T  Taxid file to set the taxonomy ids in ASN.1 deflines [File In]  Optional
  -N  Number of database volumes [Integer]  Optional
    default = 0

References

o MPIBLAST Homepage
http://www.mpiblast.org/