GZIP files are definitely our plan. I just finished testing the code that distributes the processing of a FASTQ (or pair for PE) to an arbitrary number of tasks, where each subtask extracts just the data it needs without reading any of the file it does not need. It extracts the blocks of GZIPped data into a standalone GZIP file just by copying whole blocks and appending them (if the window is not aligned perfectly, there is additional processing). Since the entire file does not need to be read, it distributes quite nicely. I'll be preparing a pull request for it soon. 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: Peter Cock [mailto:p.j.a.cock@googlemail.com] Sent: Thursday, October 06, 2011 9:19 AM To: Duddy, John Cc: Greg Von Kuster; galaxy-dev@lists.bx.psu.edu; Nate Coraor Subject: Re: [galaxy-dev] Tool shed and datatypes On Thu, Oct 6, 2011 at 5:02 PM, Duddy, John <jduddy@illumina.com> wrote:
As I understand it, Isilion is built up from "bricks" that have storage and compute power. They replicate files amongst themselves, so that for every IO request there are multiple systems that could respond. They are interconnected by an ultra fast fibre backbone.
So why not use gzipped files on top of that? Smaller chunks of data to access so should be faster even with the decompression once it gets to the CPU.
So, depending on your topology, it's possible to get a lot more throughput by working on different sections of the same file from different physical computers.
Nice.
I haven't delved into BGZF, so I can't comment. My approach to block GZIP was just to concatenate multiple GZIP files and keep a record of the offsets and sequences contained in each. The advantage is compatibility, in that it decompresses just like it was one big chunk, yet you can compose subsets of the data without decompressing/recompressing and (as long as we actually have to write out the file subsets) can reap the reduced IO benefits of smaller writes.
That sounds VERY similar to BGZF - have a read over the SAM specification which covers this. Basically they stick the block size into the gzip headers, and the BAM index files (BAI) use a 64 bit offset which is split into the BGZF block offset and the offset within that decompressed block. See: http://samtools.sourceforge.net/SAM1.pdf Peter