Keywords: DataAccWG

Hi,

Maria, Ani, Serge, and I had a small discussion about object table partitioning, here is a short summary.


Two approaches under consideration:
1) one partition per zone (a zone = "horizontal stripe"),
   clustered index on RA
2) chop zones into smaller "chunks", 1 chunk = 1 partition


Number 1 is easier to partition.
Number 2 has potential to be less I/O intensive.
Number 1 requires more data to tests.

We will test and compare both approaches.

How to test?
 - we should test with 3 full stripes.
 - height of each stripe = height of FOV
 - ok to test with smaller object density per sq deg
   (comparing to real lsst numbers)


If we use lsst numbers:
 - FOV = 3.3 x 3.3 sq deg
 - average number of objects per FOV = 4 million
 - max number of objects per FOV = 10 million
 - 2.1 K row size in Object table
we are getting:
 - 360 / 3.3 = 110 FOVs per stripe, so 330 FOVs in 3 stripes
 - 330 x  4 million objects x 2.1 K per object =  2.6 TB
 - 330 x 10 million objects x 2.1 K per object =  6.5 TB
 - plus indexes, swap space...

[my comment after the meeting: we should probably start
with ~1 million objects per FOV, which would be 650 GB of data]

Perhaps will initially use simulator provided by Sergei (LLNL)
 - need to talk to Sergei
 - need to update prototype schema

Could use small lsst-db testbed at slac (2 machines, 1TB on each)
 - will get accounts for Serge and Ani, Maria already has one


Will continue this discussion this Wednesday (db telecon) and
Friday (3 hour DataAccWG telecon)

Jacek





_______________________________________________
LSST-data mailing list
[email protected]
http://www.lsstmail.org/mailman/listinfo/lsst-data

Reply via email to