What'S Faster? "Blastall -P Blastp" Or "Blastp?"

Question

[Lee Katz]

I am just wondering if anyone has done a timed test and/or a memory test with blastall compared to blastp. In other words, classic BLAST vs BLAST+. It'd be really interesting to get some kind of frequency distribution of the times it takes to blast one set of genes vs a database using each of the tools. The comparison could include any flavor of blast.

I know it's something I can do myself, but I'm being a little lazy about it and it could be something interesting that the biostar community can get together on.

[Daniel Standage]

I haven't done the formal test you're requesting, but I have seen a steady increase in performance with each new release. Unfortunately, I've also seen bugs in the newer releases too sometimes.

Answer 1

[Simon Cockell]

This is pretty crude, but here we go. A very rapid comparison using nt and a 219 base query sequence:

$ time blastall -p blastn -i gene.fa -d /data/blastdb/nt -o out.blst  
32.19s user 3.01s system 99% cpu 35.469 total

$ time blastn -db /data/blastdb/nt -query gene.fa -out out.blstpl  
11.61s user 2.22s system 99% cpu 13.846 total

So, from this it looks like BLAST+ is somewhat faster, for a single sequence vs nt.

This is good, but I think we can do better. The BLAST+ paper reports significant speed-ups are possible when querying with long sequences. So, let's try a whole genome (even if it's only a bacterial one)...

$ time blastall -a 4 -p blastn -i NC_011353.fna -d /data/blastdb/nt -o out.blst  
46115.20s user 24.96s system 388% cpu 3:17:58.91 total

$ time blastn -num_threads 4 -db /data/blastdb/nt -query NC_011353.fna -out out.blstpl  
1462.23s user 7.98s system 233% cpu 10:29.37 total

This obviously isn't like-for-like with the first test, I ran these on 4 CPUs because I suspected they might take a while, and I was right. BLAST+ is offering a >30x speed up here. I would be interested to see if the reported reduction in memory usage is achieved too.

So it looks like for this test BLAST+ is faster (further scenarios are also offered in the paper linked above).