I did something similar, but implemented as an evolution of the original "basic" parallelism (see BWA), that: - Moved the splitting of input files into the datatype classes - Allowed any number of inputs to be split, as long as they were the same datatype (so they were mutually consistent - think paired end fastq files) - Allowed other inputs to be shared among jobs - Merged any number of outputs, which merge code implemented in the datatype classes
This worked functionally, but the IO required to split large files has proved too much for something like a whole genome (~500GB)
I was thinking of something philosophically similar to your dataset container idea, but more in the idea that a dataset is no longer a "file", so the jobs running on subsets of the dataset would just ask for the parts they need. Galaxy would take care of preserving the abstraction that the subset of the dataset is a single input file, perhaps by extracting the subset to a temporary file on local storage. Similarly, the merged outputs would just be held in the target dataset, not copied, thus making the IO cost for the "merge" 0 for the simple case where it is mere concatenation.
John Duddy Sr. Staff Software Engineer Illumina, Inc. 9885 Towne Centre Drive San Diego, CA 92121 Tel: 858-736-3584 E-mail: jduddy@illumina.com
-----Original Message----- From: galaxy-dev-bounces@lists.bx.psu.edu [mailto:galaxy-dev-bounces@lists.bx.psu.edu] On Behalf Of Andrew Straw Sent: Tuesday, August 02, 2011 7:13 AM To: galaxy-dev@bx.psu.edu Subject: [galaxy-dev] using Galaxy for map/reduce
Hi all,
I've been investigating use of Galaxy for our lab and it has many attractive aspects -- a big thank you to all involved.
We still have a couple of related sticking points, however, that I would like to get the Galaxy developers' feedback on. Basically, I want to use Galaxy to run Map/Reduce type analysis on many initial data files. What I mean is that I want to take many initial datasets (e.g. 250 or more), perhaps already stored in a library, and then apply a workflow to each and every one of them (the Map step). Then, on the many result datasets (one from each of the initial datasets), I want to run a Reduce step which creates a single dataset. I have achieved this in an imperfect and not-quite-working way with a few tricks, but I hope that with a little work, Galaxy could be much better for this type of use case.
I have a couple of specific problems and a proposal for a general solution:
1) My first specific problem is that loading many datasets (e.g. 250) into history causes the javascript running locally withing a browser to be extremely slow.
2) My second specific problem is that applying a workflow with N steps to many datasets creates even more datasets (Nx250 additional datasets). In addition to the slow Javascript problem, there seems to be other issues I haven't diagnosed further, but the console in which I'm running run.sh indicates many errors of the type "Exception AssertionError: AssertionError('State <sqlalchemy.orm.state.MutableAttrInstanceState object at 0x7f5c18c47990> is not present in this identity map',) in <bound method MutableAttrInstanceState._cleanup of <sqlalchemy.orm.state.MutableAttrInstanceState object at 0x7f5c18c47990>> ignored". Furthermore the webserver gets slow and my nginx frontend proxy gives 504 gateway time-outs.
3) There's no good way to do reduce within Galaxy. Currently I work around this by having a tool type which takes as an input a dataset and then uploads this to a self-written webserver, which then collects such uploads, performs the reduce, and offers a download link for the user to collect the reduced dataset. The user must manually then upload this dataset back into Galaxy for further processing.
My proposal for a general solution, and what I'd be interested in feedback on, is an idea of a "dataset container" (this is just a working name). It would look and act much like a dataset in the history, but would in fact be a logical construct that merely bundles together a homogeneous bunch of datasets. When a tool (or a workflow) is applied to a dataset container, Galaxy would automatically create a new container in which each dataset in this new container is the result of running the tool. (Workflows with N steps would thus generate N new containers.) The thing I like about this idea is that it preserves the ability to use tools and workflows on both individual datasets and, with some additional logic, on these new containers. In particular, I don't think the tools and workflows themselves would have to be modified. This would seemingly mitigate the slow Javascript issue by only showing a few items in the history window (even though Galaxy may have launched many jobs in the background). Furthermore, a new Reduce tool type could then act to take a dataset container as input and output a single dataset.
A library doesn't seem a good candidate for the dataset container idea I have above. I realize that a library also bundles together datasets, but it has other attributes that don't play well with the above idea (the idea of hierarchically arranged folders and heterogeneous datasets) nor can it be represented in the history.
I'm interested in thoughts on this proposal, as I think it would really help us, and I think our use case may be representative of what others might also like to do. I realize that in my text above I write "with some additional logic" to describe the work required to implement this idea, but the fact is that I have very little idea about how much work this would be. So, practically speaking, my question boils down to how hard would implementing this be, given the existing code base and goals? And, would such an implementation - if done to the taste of the Galaxy devs, of course - have a chance of making into the Galaxy distribution?
Thanks, Andrew