Experiences using cloud computing?

[thh32]

So I am currently considering using a cloud computing service as I have 180,000 blast jobs that need doing and to do that on our Uni servers would take a few months. Each job will take ~30 hours and so I was wondering what services others have used and how expensive they are etc. The main one I am looking at is Amazon as I am unaware of any others however their sales team seem to be taking ages to get back to me with pricing. Any advice would be great.

[GenoMax]

Google: https://cloud.google.com/
Microsoft: http://azure.microsoft.com/en-us/

Amazon's EC2 pricing is on the web unless you were looking for some specific discounts for your institution: http://aws.amazon.com/ec2/pricing/

What DB are you going to blast against? You probably want to use the AMI that NCBI has for Amazon to make things simple: http://blast.ncbi.nlm.nih.gov/Blast....YPE=CloudBlast

[westerman]

Given what you say -- 180,000 jobs at 30 hours each -- I suspect that Amazon will give you big thumbs up. Your Amazon instances run about $0.20/hour so a job is $6.00 and 180,000 jobs will be ... well ... more than I'd like to consider. :-)

[sarvidsson]

Given what you say -- 180,000 jobs at 30 hours each -- I suspect that Amazon will give you big thumbs up. Your Amazon instances run about $0.20/hour so a job is $6.00 and 180,000 jobs will be ... well ... more than I'd like to consider. :-)

I'd start thinking about alternative ways to perform that analysis... You'd be busy for months just handling the logistics of running these jobs.

[thh32]

We are currently blasting against Swiss prot and the trembl sections specific for bactieria and archaea. Also yes I was hoping I could disciss prices with them but the free 750 hours you get per month could be quite useful as I hadnt seen that before.

[GenoMax]

Perhaps you are not running your blast jobs efficiently? Just a thought.

Swissprot/Trembl restricted to bacteria has got to be smaller than nr. 30 hours seems fairly long unless your input file has hundreds/thousands of sequences.

[sarvidsson]

I don't know what you are BLASTing, but if it is partly redundant, you may want to remove redundancy before running the jobs...

[thh32]

Yes this is my issue, just looked at the cost and its going to be better to buy a whole load of new nodes for our Uni HPC instead.

[thh32]

Perhaps you are not running your blast jobs efficiently? Just a thought.

Swissprot/Trembl restricted to bacteria has got to be smaller than nr. 30 hours seems fairly long unless your input file has hundreds/thousands of sequences.

Each of the query files is ~9Mb as the original file of 6.5Gb was split into 1000 smaller pieces to speed up the process however the bacteria subset of Trembl is 30Gb which seems to be the issue but even when split into 1Gb subsets it still haves 30 hours. How do you increase the efficiency of your blast jobs?

[GenoMax]

Can you elaborate what exactly you are trying to do with the blasting? You are using multiple threads for the blast?

[westerman]

One problem of running multiple Blast jobs on a cluster is reading in the Blast database into each cluster node. I find that if I run Blast on too many nodes, even with a screaming fast file server, my I/O wait time goes sky high.

Another possible solution is to use the program called 'Diamond' which is a blastx replacement.

[thh32]

Can you elaborate what exactly you are trying to do with the blasting? You are using multiple threads for the blast?

We are trying to provide functional annotation to an assembly we have recently created. We are using 1 core per blast job as to allow as many as possible to get onto the server at once.

[GenoMax]

We are trying to provide functional annotation to an assembly we have recently created. We are using 1 core per blast job as to allow as many as possible to get onto the server at once.

As Rick mentioned above that is probably not good since each of those jobs is trying to read the 30G database simultaneously on the same node.

Try using all cores on a physical server for one job with multiple threads (depending on the scheduler you should be able to ask it to run those threads on one physical machine) and see if that speeds things up. Logically it should, though I can't predict the drop in number from 30h per job (since you would still need to chunk through an equivalent number of jobs sequentially).

If you have access to a server with enough RAM you could try making a RAMdisk, cache the database there and do without disk access for index access part. Worth a try.

[westerman]

Putting everything into memory is a good. Ramdisk or just letting blast run in a large memory space. AWS has some large memory multi-cpu machines -- 60 GB upward -- which would allow for a test of the concept.

[mbblack]

Depending on what you are looking for and your stringency requirements, could you switch to BLAT instead? Maybe adopt a tiered approach of a first pass with BLAT to reduce the search space, then BLAST or HMMER (in parallel runs) for the higher stringency search on selected hits.