Thank you Jeff.  I am sinking my teeth into this.
I will likely call upon my Data Products WG colleagues
for inputs along the way.

- Kirk

----- Original Message -----
From: Jeffrey P Kantor <[EMAIL PROTECTED]>
Date: Wednesday, June 7, 2006 11:57 am
Subject: Re: [LSST-data] Growth of database size

> Hello all,
> 
> We have 2 major sources of requirements for disk I/o:
> - Pipelines (read/write)
> - End user access/queries (mostly read)
> 
> We understand the first pretty well, but getting a handle on the 
> second has
> been a much bigger problem.  So far, we have the following completely
> independent models of this access:
> 
> - SDSS history and Sergei's extrapolation of it to queries
> - Kem's 7-page list of science examples
> - Tim's list of representative Level 3 data products in the Functional
> Requirements Spec
> - The UML model Community Science use cases
> 
> I have asked Kirk Borne to take the lead in examining these items 
> and come
> up with the "baseline" list of use cases and queries that we will 
> use to do
> sizing for this.  It is due at the end of June.  I have asked 
> Chris Smith to
> assist Kirk in this endeavor.  I would ask the rest of the Data 
> ProductsWorking Group to offer assistance as well.
> 
> This is a critical hole in our cost estimating that makes me 
> increasinglynervous.  I hope that we will have a good model of 
> this by the end of the
> month.
> 
> Jeff
> 
> 
> > From: Jacek Becla <[EMAIL PROTECTED]>
> > Organization: Stanford Linear Accelerator Center
> > Reply-To: LSST Data Management <[email protected]>
> > Date: Tue, 06 Jun 2006 15:42:06 -0700
> > To: Jim Gray <[EMAIL PROTECTED]>
> > Cc: LSST Data Management <[email protected]>
> > Subject: Re: [LSST-data] Growth of database size
> > 
> > Hi Jim,
> > 
> > 
> > I completely agree with you that the biggest problem will be 
> disk I/O
> > and that we should not underestimate it. The only reason why I 
> have not
> > done the estimates yet is that the input data that I need for such
> > exercise is pretty much unknown. To estimate disk IO that a database
> > will generate/need, we need to know things like:
> > - estimated number of queries
> > - types of common queries
> > - estimated size of data that these queries will need
> > - indexes available and type of indexes
> > - how data will be clustered
> > - how data will be partitioned
> > - access patterns
> > - database engine used
> > - and others...
> > 
> > As far as I know we do not have the answers or good estimates 
> for any of
> > the issues listed above, so I would not trust the disk IO 
> numbers that
> > we could come up with today. If you have any suggestions how to
> > realistically estimate disk IO for LSST today, I'd love to talk 
> to you!
> > 
> > 
> > Thanks,
> > Jacek
> > 
> > p.s. We could do the disk IO estimate for Data Ingest, but as
> > we both know the IO there is not very challenging (assuming we
> > partition indexes correctly)
> > 
> > 
> > 
> > 
> > 
> > Jim Gray wrote:
> >> LSST sizing should focus more on disk IO/s than on disk bytes.
> >> 
> >> The database size is definitely an issue -- it tells you how 
> much disk
> >> capacity and how much network bandwidth you need and if you 
> know the
> >> instruction density (instructions per object it implies the cpu 
> demand).>> 
> >> So, DB size is VERY good to know.
> >>  
> >> But... In the 2012 world of 10GB/s network links, 10TB disks, 
> and 100
> >> core processors, these are not the scarce resources (we hope).
> >> Indeed, you can afford a 10x cpu cost converting to and from 
> ASCI/csv>> rather than using binary for data ingest just to have a 
> simpler>> interface in the pipeline.
> >> 
> >> As far as I can tell the scarce computing resource will be disk IO.
> >> 
> >> Now we are expecting about 12 TB/night.
> >> In rough numbers, one disk worth of data per night.
> >> Each of the 10TB disks will deliver about 250 IO/s for small random
> >> requests,  So if we triplex the disks we get about 3x200 = 750 
> IO/s to
> >> do the processing.
> >> The disk bandwidth goes as the square root of the aerial 
> density so we
> >> can expect about 4x more bandwidth or 250MBps.
> >> 
> >> 1e13 nightly LSST bytes written at 2.5e8 bytes/sec is 4e4 
> seconds or 10
> >> hours 
> >> -- so the disks can be written in 10 hours but there is not a 
> lot of
> >> slack to read them.
> >> These disks will deliver about 250 TINY random Ios/s.
> >> If you do LARGE 1MB reads and writes then the transfer time is
> >> significant and the number drops to 125 IO/s
> >> 
> >> page size (B)    seek time (ms)    transfer time (ms)    random
> >> transfers/sec
> >> 1,000             4.00              0.00                    249.75
> >> 10,000            4.00              0.04                    247.52
> >> 100,000           4.00              0.40                    227.27
> >> 1,000,000         4.00              4.00                    125.00
> >> 10,000,000        4.00              40.00                    22.73
> >> 100,000,000       4.00              400.00                    2.48
> >> 
> >> And significantly, (again for random Ios)
> >> page size (B)    Bandwidth (MB/s)
> >> 1,000                0
> >> 10,000               2
> >> 100,000             23
> >> 1,000,000          125
> >> 10,000,000         227
> >> 100,000,000        248
> >> 
> >> So, you should count on the DBMS doing 1MB/s and giving you 125
> >> MBps/disk and using MASSIVE main memory (this is a page size 
> 100x bigger
> >> than today's sizes).
> >> Now you are back to needing lots more disks/night or designing 
> the disk
> >> arrays to use all the arms all the time.
> >> 
> >> It is ESSENTIAL that the LSST pay attention to the disk IO/s issue.
> >> It will be a gating technology (capacity will not be).
> >> The simple way to think of this is just to imagine that each 
> disk has
> >> infinite capacity but delivers only 150 IO/s.
> >> The LSST IO/S requirements will imply WAY more disk capacity 
> than will
> >> be needed to just store the data and indices.
> >> 
> >> So these discussions of "database size" are great,
> >> But they should all include the IO bandwidth (MB/s) and IO per 
> second>> requirements. 
> >> 
> >> I attach the spreadsheet if you want to try different 
> parameters on this
> >> simple model
> >> 
> >> 
> >> Jim Gray
> >> Microsoft Research,  Suite 1690, 455 Market, SF CA 94105, tel: 
> 415 778
> >> 8222 fax: 425 706 7329 [EMAIL PROTECTED]
> >> http://research.Microsoft.com/~gray
> >> 
> >> 
> >> -----Original Message-----
> >> From: [EMAIL PROTECTED]
> >> [mailto:[EMAIL PROTECTED] On Behalf Of Kem Cook
> >> Sent: Monday, June 05, 2006 11:55 PM
> >> To: [EMAIL PROTECTED]; LSST Data Management
> >> Subject: Re: [LSST-data] Growth of database size
> >> 
> >> Hi All,
> >> 
> >> I agree with Tim's esimates, but there are details which 
> haven't been
> >> fleshed out.  There are parameters which don't really add 
> volume to the
> >> data, but they are there.  The time dependent database needs  
> motion>> information: parallax, proper motion or orbital 
> parameters.  The time
> >> dependent objects will also contain added information in terms 
> of the
> >> likelihood of blendedness, multiplicity and variability parameters.
> >> These data are added on a per object basis and as such, do not
> >> significantly increase the volume of data, but should not be 
> forgotten.>> Presumably, these parameters will be present from the 
> first detection of
> >> a time dependent object and will not increase in volume with time.
> >> 
> >> Kem
> >> 
> >> 
> >>> Hi Jacek,
> >>> 
> >>> I have created a simple model for how the size of the object 
> database>>> will grow between data releases (DR).   Here are my 
> assumptions:>>> 
> >>> 1.  Data releases occur every 6 months
> >>> 
> >>> 2.  We meet our SRD requirements of 100 visits per field per year
> >>> 
> >>> 3.  The database is split into two parts.   The first, 
> dominated by
> >>> galaxies, contains the static information for every object 
> detected at
> >> 
> >> 
> >>> that point in the survey, mostly generated by combining the
> >> 
> >> information
> >> 
> >>> in image stacks.   I'll call this the 'deep database'   The 
> second,>>> dominated by stars, contains the time dependent 
> information for
> >> 
> >> objects
> >> 
> >>> bright enough to be usefully detected in individual exposures. 
>  I'll
> >>> call this the 'time dependent database'.
> >>> 
> >>> 4.  An object record in the deep database is about 100 bytes:  
> 6 band
> >>> magnitude + errors; data quality flags; shape information.
> >>> 
> >>> 5.  An object record in the time dependent database is about 
> 10 bytes:
> >>> 1 band magnitude + error + data quality flags.
> >>> 
> >>> 6.  For the first DR, the limiting magnitude for the time 
> dependent>>> database is 24.5 (where it remains), while the 
> limiting magnitude for
> >>> the deep database is already at about 26.1 from stacking 20 R band
> >>> images.   So at DR1, there are already about 20 times more 
> objects in
> >>> the deep database than in the time dependent.
> >>> 
> >>> Consider first the growth of the deep database.   The limiting 
> flux to
> >>> fixed signal-to-noise will decrease as 1/sqrt(n_exp), where 
> n_exp is
> >> 
> >> the
> >> 
> >>> number of exposures effectively stacked and used for 
> detection.   I
> >>> assume that measurement occurs in all bands, but detection 
> occurs only
> >>> in the R band.   The SRD calls for 40 R band exposures per 
> field per
> >>> year, or 20 additional for every DR.    The limiting magnitude
> >> 
> >> increases
> >> 
> >>> as 1.25*log (20DR), and we go progressively fainter in the galaxy
> >>> brightness distribution.   I've taken the galaxy data here 
> from the
> >>> Subaru Deep Field, which gives the slope of the cumulative 
> brightness>>> distribution to be d(logN)/d(mag) = 0.45 in the 
> region of interest.
> >>> The size of the deep database then grows as 100 * 
> (20DR)**(0.45 *
> >>> 1.25)
> >>> 
> >>> The time dependent database grows strictly linearly with the 
> number of
> >> 
> >> 
> >>> observations in all bands, which is 50 per DR, so it goes as 
> 10 * (50
> >> 
> >> DR).
> >> 
> >>> Taking account of the factor of 20 difference in number of 
> objects at
> >>> DR1,  two attached plots show the growth of the deep database 
> size,>>> and the growth of both together.  The roughly square root 
> growth of
> >>> the deep data dominates the first half of the survey, but is then
> >>> overtaken in the second half by the linear growth of the time
> >> 
> >> dependent database.
> >> 
> >>> In spite of my many assumptions, which are unlikely to be 
> right in
> >>> detail, I think the overall behavior is about right.
> >>> 
> >>> Let me know if you see an error or need more information.
> >>> 
> >>> Cheers,
> >>> Tim
> >>> 
> >>> 
> >>> _______________________________________________
> >>> LSST-data mailing list
> >>> [email protected]
> >>> http://www.lsstmail.org/mailman/listinfo/lsst-data
> >>> 
> >> 
> >> 
> >> _______________________________________________
> >> LSST-data mailing list
> >> [email protected]
> >> http://www.lsstmail.org/mailman/listinfo/lsst-data
> >> 
> >> 
> >> ----------------------------------------------------------------
> --------
> >> 
> >> _______________________________________________
> >> LSST-data mailing list
> >> [email protected]
> >> http://www.lsstmail.org/mailman/listinfo/lsst-data
> > 
> > _______________________________________________
> > LSST-data mailing list
> > [email protected]
> > http://www.lsstmail.org/mailman/listinfo/lsst-data
> > 
> 
> _______________________________________________
> LSST-data mailing list
> [email protected]
> http://www.lsstmail.org/mailman/listinfo/lsst-data
> 
_______________________________________________
LSST-data mailing list
[email protected]
http://www.lsstmail.org/mailman/listinfo/lsst-data

Reply via email to