Downloading gzip'd fastq

fq_download.md

Source	Dst. file type	Protocol	Time (s)	Command Line
NCBI	.sra	ftp	296	wget
NCBI	.fastq.gz	sra toolkit	~23000	fastq-dump -Z --gzip --split-spot
local file	sra=>fastq.gz	sra toolkit	~15000	fastq-dump --gzip --split-spot --split-3
EBI	.fastq.gz	aspera	513+492	aspera -QT -l 300m
EBI	.fastq.gz	ftp	1876+1946	wget

Notes:

• Destination: a super computer at Broad Institute (yes, the connection is that fast – 22GB in 5 minutes via ftp).

• 513+492 means downloading read1 took 513 wall-clock seconds and read2 492 seconds. The two files were downloaded in parallel.

• Time on downloading and file conversion with SRA toolkit is estimated based on partially downloaded/converted file sizes. Inaccurate.

• SRA toolkit may be spending significant time on gzip compression. It is probably faster to convert to plain FASTQs first and then run gzip afterwards.

URLs:

# SRA ftp, SRA format
ftp://ftp-trace.ncbi.nih.gov/sra/sra-instant/reads/ByRun/sra/SRR/SRR208/SRR2088062/SRR2088062.sra
# ENA ftp, gzip'd FASTQ format
ftp://ftp.sra.ebi.ac.uk/vol1/fastq/SRR208/002/SRR2088062/SRR2088062_1.fastq.gz
ftp://ftp.sra.ebi.ac.uk/vol1/fastq/SRR208/002/SRR2088062/SRR2088062_2.fastq.gz
# ENA aspera, gzip'd FASTQ format
era-fasp@fasp.sra.ebi.ac.uk:/vol1/fastq/SRR208/002/SRR2088062/SRR2088062_1.fastq.gz
era-fasp@fasp.sra.ebi.ac.uk:/vol1/fastq/SRR208/002/SRR2088062/SRR2088062_2.fastq.gz

There are significant differences in the performance of fastq-dump, whether you ask for gzip output (3-4x slower), write to a file with --split-3 instead of stdout (1.5x slower).

Gzip	To stdout	Command line	wall-clock time
x	✓	fastq-dump -Z --split-spot	228s
x	x	fastq-dump --split-spot --split-3	352s
✓	✓	fastq-dump -Z --gzip --split-spot	1060s
✓	x	fastq-dump --gzip --split-spot --split-3	1155s

Tested on a local file, SRR1028232.sra (1 GB), with sratoolkit version 2.5.7 centos binaries. On a server in Pennsylvania. Time to download the SRA file using prefetch: 90 seconds.

In the post above, local .sra=>.fastq.gz is 50x slower than downloading the .sra. In this experiment here, it is 13x slower. Still, I think fastq-dump would need to be faster.

Thanks a lot, @rchikhi. Just saw your comment. I was using --gzip because most of time we would not want to keep plain fastq. On an additional note, in my table, running fastq-dump on remote SRA accessions seems much slower than wget download + local fastq-dump. Have you observed this? If this is true, probably NCBI should not hide the FTP download links to SRA files.