[Beowulf] Station wagon full of tapes
Robert G. Brown
rgb at phy.duke.edu
Tue May 26 08:16:57 PDT 2009
On Tue, 26 May 2009, Chris Dagdigian wrote:
> I deal quite often with the "next-gen" DNA sequencing instruments that
> produce 1TB/day in TIFF images that are then distilled down to the DNA
> basecalls before the short reads are subjected to alignment. Then the
> resulting longer sequences are usually aligned again against a reference
> Lots of data, lots of computation.
> The 1 Terabyte of TIFF images typically reduces down to about 200 GB in
> intermediate data which is further distilled down into a few hundred KB of
> actual sequence data. The entire process is interesting and it is a massive
> Bio/IT challenge as these types of terabyte-scale data producing lab
> instruments are popping up everywhere (the cost of one of these instruments
> is now easily within reach of a single grant-funded researcher at a facility
> of any size...). We are only a few technology revolutions away from these
> boxes showing up in your point of care primary physician's office (well not
> really, probably a backend service lab that your physician outsources to ...)
> Anyway the new data ingestion service that Amazon offers is, I think, going
> to be a big deal in our field.
Sure, but why wouldn't it be cheaper for e.g. NSF or NIH to fund an
exact clone of the service Amazon plans to offer and provide it for free
to its supported research groups (or rather, do bookkeeping but it is
all internal bookkeeping, moving money from one pocket to another).
Amazon has to make a profit. Granting agencies don't have to pay the
profit that Amazon has to make. Amazon has to take substantial risks to
make its profit. Granting agencies have no risk.
All of the things you assert for DNA sequencing are true for high energy
physics. Enormous datasets, lots of computation. HEP's INTERNATIONAL
solution is ATLAS, not Amazon.
Supporting commercial access into such a DB a la >>google<< but for
genomic data, sure, but that's not really cluster computing, that's a
large shared DB. I could see that as a spin off data service of Amazon
or Google or a new business altogether, but I'd view it as a niche and
not really HPC.
Grant funded research involving large scale shared data resources can
ALWAYS be done more cheaply than by buying the data services from
profit-making third parties unless there are nonlinear e.g. proprietary
IP barriers. This is trebly true given that research facilities are
typically on a very high speed networks e.g. lambda rail that the
government is funding anyway, where Amazon or other commercial third
parties have to rent time on those networks and then resell the rental
back to the government at a profit or use slower commercial networks and
with the same sort of throughput markup.
Are there any such barriers here? I'd have to say that I would be most
unhappy seeing my own tax dollars going to make Amazon shareholders rich
when they could be spent more efficiently without a middleman raking in
a 50 to 100% markup on the service. Of course I'm easily irked -- when
I think of all the money spent on Windows by the US government it makes
my blood boil.
I'd want to see a solid CBA proving that this is the cheapest way to
proceed before dumping tons of tax money into it, if I were king of the
world (or just in charge of a major granting agency).
> For the following reasons:
> - Bio people are being buried in data
> - Once we process the data to get the derived results, the primary data just
> needs to go somewhere cheap
> - Amazon and other internet-scale people can do peta-scale or exa-scale
> storage far better & cheaper than any of my customers
> - These instruments are popping up in wet labs across campus with weak/anemic
> network links to IT core facilities and data centers
> - Scientists in many cases are required to share data that is grant funded
> - Amazon has some neat "downloader pays" models that make it easier for
> researchers to affordably offer up peta-scale data sets for sharing
> I suspect that very large amount of scientific data will be making a 1-way
> trip into the cloud. The data will stay there "forever" as a deep store. In
> the ocasional cases where the data needs to be re-processed or re-analyized
> it would be not unreasonable to fire up some cloud server nodes to do the
> re-work in-situ.
> The disk ingest service was the final piece. I can see this happening in life
> science environments:
> - Massive data generated in the wet lab
> - Captured to local storage (10 - 40TB) with small HPC component
> - Data is processed locally into derived and distilled forms
> - Derived data replicated to campus/lab facilities for online primary storage
> - Derived data (and possibly the full raw data) is compressed, placed onto
> drives and ingested into Amazon for long term storage
> - If re-analysis is ever needed, have existing EC2 AMIs preloaded with the
> necessary software
> Basically it comes down to the fact that Amazon may be able to offer
> big-yet-slow storage in the terabyte to petabyte range at levels of cost and
> geographical redundancy that would be extremely difficult to match with local
> resources at a small non-specialized organization.
> My $.02 of course
> On May 26, 2009, at 8:58 AM, Jeff Layton wrote:
>> Gerry Creager wrote:
>>> There was an interesting brainstorming session at Rocks-A-Palooza a couple
>>> of weeks ago. Someone wants to offer Amazon resources. Problem remains
>>> for me: How can I get sufficient cloud resources for computing (I'll
>>> hammer on dataset transport in a moment) that will handle reasonable
>>> weather models with their small message MPI chatter, and lots of file I/O?
>>> I've been assured that Amazon's ready to accommodate that.
>> This is one of the problems - clouds aren't ready for this kind of
>> usage model yet. They only have GigE and usually it's oversubscribed.
>> When you say file IO, they hear capacity, not performance (either
>> throughput or IOPS). And as you point out, the pipe to/from the
>> cloud is not ready for lots of data.
>>> However, getting data into S3 for availability, when a daily
>>> multi-gigabyte dataset is used for initiation, and another is created as
>>> output, is going to be expensive, and likely slow. I think there are
>>> other approaches that have to be evaluated. I am not sure the cloud is
>>> ready for MPI play on a significant basis, just yet.
>> I haven't seen the cloud ready yet for anything other than embarrassingly
>> parallel codes (i.e. since node, small IO requirements). Has anyone seen
>> differently? (as an example of what might work, CloudBurst seems to be
>> gaining some traction - doing sequencing in the cloud. The only problem
>> is that sequencing can generate a great deal of data pretty rapidly).
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
>> To change your subscription (digest mode or unsubscribe) visit
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit
Robert G. Brown http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567 Fax: 919-660-2525 email:rgb at phy.duke.edu
More information about the Beowulf