SAND User's Manual

Last Updated January 2011

Overview

SAND is a set of modules for accelerating genome assembly and other bioinformatics tasks. Using the Work Queue framework, SAND can distribute computational tasks to hundreds of nodes harnessed from clusters, clouds, grids, or just the idle machines you have in your office. SAND can be used as a drop-in replacement for the conventional overlapper in the Celera Assembler, or can be used as a standalone tool by advanced users.

SAND is part of the Cooperating Computing Tools. You can download the CCTools from this web page, follow the installation instructions, and you are ready to go.

Using SAND with the Celera Assembler

If you are already using the Celera Assembler version 5.4, you can easily switch to using SAND to accelerate assemblies by using our modified sand_runCA_5.4 script. If you are using the newer Celera Assembler version 6.1, you can switch to SAND by using the sand_runCA_6.1 script. If you are using Celera Assembler version 7.0, you can switch to SAND by using the sand_runCA_7.0 script. We assume that you have installed SAND according to the instructions above and addeed it to your PATH.

Copy the script sand_runCA_X.X into the same directory where you have the normal runCA installed.
Set ovlOverlapper=sand to your spec file.
For Celera 7.0, add sandAlignFlags=-e "-o ovl_new" to your spec file. This tells the alignment program to output new Celera 7.0 style overlaps.
Run sand_runCA_X.X just like you normally use runCA

You will see the assembly start as normal. When the overlapping stage begins, you will see output like this:

 Total | Workers   | Tasks                      Avg | Candidates
  Time | Idle Busy | Submit Idle  Run   Done   Time | Found
     0 |    0    0 |      0    0    0      0    nan | 0
     5 |    0    0 |      0    0    0      0    nan | 0
    10 |    0    0 |      0    0    0      0    nan | 0
    15 |    0    0 |      0    0    0      0    nan | 0
...

SAND is now waiting for you to start work_queue_worker processes. Each worker that you start will connect back to the master process nd perform small pieces of the work at a time. In general, the more machines that you can harness, the faster the work will go.

For testing SAND, just open a new console window and start a single worker, specifying the hostname where the master runs and the port number it is listening on. (This can be changed with the sandPort option in the CA spec file.) For example:

% work_queue_worker master.somewhere.edu 9123

With one worker, you will see some change in the output, like this:

 Total | Workers   | Tasks                      Avg | Candidates
  Time | Idle Busy | Submit Idle  Run   Done   Time | Found
     0 |    0    0 |      0    0    0      0    nan | 0
     5 |    0    1 |    100   83    1     16   0.32 | 1858
    10 |    0    1 |    100   69    1     30   0.33 | 3649
    15 |    0    1 |    100   55    1     44   0.34 | 5464

For a very small assembly, a single worker might be sufficient. However, to run a really large assembly at scale, you will need to run as many workers as possible. A simple (but tiresome) way of doing so is to ssh into lots of machines and manually run work_queue_worker as above. But, if you have access to a batch system like Condor or SGE, you can use them to start many workers with a single submit command.

We have provided some simple scripts to make this easy. For example, to submit 10 workers to your local Condor pool:

% condor_submit_workers master.somewhere.edu 9123 10
Submitting job(s)..........
Logging submit event(s)..........
10 job(s) submitted to cluster 298.

Or, to submit 10 worker processes to your SGE cluster:

% sge_submit_workers master.somewhere.edu 9123 10
Your job 1054781 ("worker.sh") has been submitted
Your job 1054782 ("worker.sh") has been submitted
Your job 1054783 ("worker.sh") has been submitted
...

Note that condor_submit_workers and sge_submit_workers are simple shell scripts, so you can edit them directly if you would like to change batch options or other details.

Once the workers begin running, the SAND modules can dispatch tasks to each one very quickly. It's ok if a machine running a worker crashes or is turned off; the work will be silently sent elsewhere to run.

When the SAND module's master process completes, your workers will still be available, so you can either run another master with the same workers, remove them from the batch system, or wait for them to expire. If you do nothing for 15 minutes, they will automatically exit.

SAND in More Detail

This section explains the two SAND modules in more detail, if you would like to tune the performance or use them independently of Celera. We assume that you begin with a file of reads in FASTA format. To use the SAND modules, you must first generate repeats and compress the data. (If you don't have data handy, download small.cfa and small.repeats data from the SAND Webpage.) To generate repeats from a FASTA file, use the meryl tool from the Celera Assembler:

% meryl -B -m 24 -C -L 100 -v -o small.meryl -s small.fa
% meryl -Dt -s small.meryl -n 100 > small.repeats

Then use sand_compress_reads to compress the sequence data into a compressed FASTA (.cfa) file:

% sand_compress_reads small.fa small.cfa

The filtering step will read in the compressed sequence data (small.cfa) and quickly produce a list of candidate sequences (small.cand) for the following step to consider in detail. Start the filtering step as follows:

% sand_filter_master -r small.repeats small.cfa small.cand

While the filtering step runs, it will print some statistics to the console, showing the number of workers available, tasks running, and so forth:

 Total | Workers   | Tasks                      Avg | Candidates
  Time | Idle Busy | Submit Idle  Run   Done   Time | Found
     0 |    0    0 |      0    0    0      0    nan | 0
     5 |    0   14 |    158   14   14    130   0.39 | 16452
    10 |    0   15 |    356   15   15    326   0.38 | 42382
    15 |    0   15 |    549   15   15    519   0.38 | 69055
    20 |    0   15 |    744   15   15    714   0.38 | 96298
    25 |    0   15 |    942   15   15    912   0.38 | 124284

The alignment step will take the list of candidates generated in the previous step (small.cand), the compressed sequences (small.cfa) and produce a listing of how and where the sequences overlap (small.ovl). For example:

% sand_align_master sand_align_kernel -e "-q 0.04 -m 40" small.cand small.cfa small.ovl

The options -q 0.04 -m 40 passed to sand_align_kernel indicate a minimum alignment quality of 0.04 and a minimum alignment length of 40 bases. Again, a progress table will be printed to standard out:

 Total | Workers   | Tasks                      Avg | K-Cand K-Seqs | Total
  Time | Idle Busy | Submit Idle  Run   Done   Time | Loaded Loaded | Speedup
     0 |    0    0 |      0    0    0      0   0.00 |      0      0 |  0.00
     8 |    0   48 |    100   52   48      0   0.00 |   1000    284 |  0.00
    10 |    0   86 |    100   13   86      1   7.07 |   1000    284 |  0.71
    36 |    1   83 |    181   14   83      2  19.47 |   1810    413 |  1.08
   179 |    1   83 |    259   92   83      3  22.51 |   2590   1499 |  0.38
   186 |    2   80 |    259   15   80     85  28.54 |   2590   1499 | 13.04
   199 |    2   80 |    334   90   80     86  29.96 |   3340   1499 | 12.95
   200 |    2   80 |    334   90   80    114  59.43 |   3340   1499 | 33.88
   202 |    2   81 |    334    9   81    165  86.08 |   3340   1499 | 70.32

After the sequence alignment step completes, you will have an overlap (.ovl) file that can be fed into the final stages of your assembler to complete the consensus step.

Tuning Suggestions

As a rule of thumb, a single task should take a minute or two. If tasks are much longer than that, it becomes more difficult to measure progress and recover from failures. If tasks are much shorter than that, the overhead of managing the tasks becomes excessive. Use the -n parameter to increase or decrease the size of tasks.

When using banded alignment (the default), the -q match quality parameter has a significant effect on speed. A higher quality threshhold will consider more alignments, but take longer and produce more output.

The columns of the output are as follows:

Total Time is the elapsed time the master has been running.

Workers Idle is the number of workers that are connected, but do not have a task to run.

Workers Busy is the number of workers that are currently running a task.

Tasks Submitted is the cumulative number of tasks created by the master.

Tasks Idle is the number of tasks waiting for a worker.

Tasks Running is the number of tasks currently running on a worker.

Tasks Done is the cumulative number of tasks completed.

Avg Time is the average time a task takes to run. An average time of 60 seconds is a good goal.

K-Cand Loaded indicates the number of candidates loaded into memory (in thousands).

K-Seqs Loaded indicates the number of sequences loaded into memory (in thousands).

Speedup is the approximate speed of the distributed framework, relative to one processor.

For More Information

For the latest information about SAND, please visit our web site and subscribe to our mailing list.