Data Exercise 1.3: Splitting Large Input for Better Throughput¶

The objective of this exercise is to prepare for blasting a much larger input query file by splitting the input for greater throughput and lower memory and disk requirements. Splitting the input will also mean that we don't have to rely on additional large-data measures for the input query files.

Setup¶

1. Log in to login04.osgconnect.net

2. Create a directory for this exercise named blast-split and change into it.

3. Copy over the following files from the previous exercise:
   • Your submit file
   • blastx
   • pdbaa_files.tar.gz
   • blast_wrapper.sh
4. Remember to modify the submit file for the new locations of the above files.

Obtain the large input¶

We've previously used blastx to analyze a relatively small input file of test data, mouse.fa, but let's imagine that you now need to blast a much larger dataset for your research. This dataset can be downloaded with the following command:

user@login04 $ wget http://stash.osgconnect.net/public/osgvsp20/mouse_rna.tar.gz

After un-tar'ing the file, you should be able to confirm that it's size is roughly 100 MB. Not only is this near the size cutoff for HTCondor file transfer, it would take hours to complete a single blastx analysis for it and the resulting output file would be huge.

Split the input file¶

For blast, it's scientifically valid to split up the input query file, analyze the pieces, and then put the results back together at the end! On the other hand, BLAST databases should not be split, because the blast output includes a score value for each sequence that is calculated relative to the entire length of the database.

Because genetic sequence data is used heavily across the life sciences, there are also tools for splitting up the data into smaller files. One of these is called genome tools, and you can download a package of precompiled binaries (just like BLAST) using the following command:

user@login04 $ wget http://stash.osgconnect.net/public/osgvsp20/gt-1.5.10-Linux_x86_64-64bit-complete.tar.gz

Un-tar the gt package (tar -xzvf ...), then run its sequence file splitter as follows, with the target file size of 1MB:

user@login04 $ ./gt-1.5.10-Linux_x86_64-64bit-complete/bin/gt splitfasta -targetsize 1 mouse_rna.fa

You'll notice that the result is a set of 100 files, all about the size of 1 MB, and numbered 1 through 100.

Run a Jobs on Split Input¶

Now, you'll submit jobs on the split input files, where each job will use a different piece of the large original input file.

Modify the submit file¶

First, you'll create a new submit file that passes the input filename as an argument and use a list of applicable filenames. Follow the below steps:

1. Copy the submit file from the previous exercise to a new file called blast_split.sub and modify the "queue" line of the submit file to the following:

   queue inputfile matching mouse_rna.fa.*
   

2. Replace the mouse.fa instances in the submit file with $(inputfile), and rename the output, log, and error files to use the same inputfile variable:

   output = $(inputfile).out
   error = $(inputfile).err
   log = $(inputfile).log
   

3. Add an arguments line to the submit file so it will pass the name of the input file to the wrapper script

   arguments = $(inputfile)
   

4. Add the $(inputfile) to the end of your list of transfer_input_files:

   transfer_input_files = ... , $(inputfile)
   

5. Update the memory and disk requests, since the new input file is larger and will also produce larger output. It may be best to overestimate to something like 1 GB for each.

Modify the wrapper file¶

Replace instances of the input file name in the blast_wrapper.sh script so that it will insert the first argument in place of the input filename, like so:

./blastx -db pdbaa -query $1 -out $1.result

Submit the jobs¶

This job will take a bit longer than the job in the last exercise, since the input file is larger (by about 3-fold). Again, make sure that only the desired output, error, and result files come back at the end of the job. In our tests, the jobs ran for ~15 minutes.

user@login04 $ rm *.error *.log *.out *.result

Update the resource requests¶

After the job finishes successfully, examine the log file for memory and disk usage, and update the requests in the submit file. In Exercise 2.1 you'll submit many jobs at once and use a different method for handling the pdbaa_files.tar.gz file, which is a bit too large to use regular file transfer when submitting many jobs.