Hello, If we assume we have 2.3 TB of RAM and use it for common indexes, (and force the queries to use these indexes even if selectivity is high), then the number of disks needed = 4,400.
DISKS, not disk heads(!). At the last telecon we (incorrectly) assumed that this number is for disk heads. I spoke with our local experts who confirmed that the number really is for "disk head assemblies" (or spindles) because all disk heads move together and there is one disk head assembly per spindle (at least in most drives available today). I also checked with SLAC folks about number of disk spindles to support BaBar, and at the moment it is ~2,000. So the 1,000 disks for LSST proposed by the top-down estimate seems rather low comparing to BaBar numbers. 4,400 seems more realistic. The number of disks is highly dependent on data block size, e.g.: data number block of size disks [KB] --------------- 16 43,200 32 22,100 64 11,700 128 6,800 <-- I'd suggest to use either this 256 4,400 <-- or this 512 3,400 1024 3,300 The small-ish disks ~2013 will probably be ~0.5 TB in size, which would give us 2 PB of disk storage assuming 4,400 disks. That sounds like a very good fit to what the database size estimates suggest: 1.6 TB of disk space for DR2. If we want to be less aggressive, we can pick the 128K data block size. Assuming 12 disks per array, 2 unit high, 40 rack units per rack, and back-to-back organization, we end up with 480 disks per rack, and only 9 racks needed. The spreadsheet that supports these numbers: http://www.slac.stanford.edu/~becla/tmp/lsst_diskIO_estimates_v07.xls (I am still working on documentation and cleaning up the spreadsheet.) BTW, all these numbers are for the usable disk space, we should not forget about the RAID overheads, and disk spares. Jacek _______________________________________________ LSST-data mailing list [email protected] http://www.lsstmail.org/mailman/listinfo/lsst-data
