Hoi,
modulo any errors in my computations, the production 2004 for 150/fb
data and 500/fb BB MC will need
1.3TB disk space (+/- 150GB)
O(60000) CPU hours (assuming that the tagbits are functional in SP5)
O(150000) CPU hours (worst case)
Deleting the moments-analysis (and other) hbook files will free up
O(750GB) of AWG diskspace, i.e., we'd need to find about 600GB.
Neither scenario is a serious problem to achieve in two months
(barring catastrophic SCS/batch breakdowns). See below for details.
Seems like a good deal to get two theses out by fall and produce
publication-ready results before BELLE publishes their combination of
b->sg and b->ulnu as another "first" ...
Cheers,
--U.
----------------------------------------------------------------------
1. Current ISL disks
====================
We have the following AWG disks:
shire01>df -k /nfs/farm/babar/AWG37/. /nfs/farm/babar/AWG23/. /nfs/farm/babar/AWG18/.
Filesystem kbytes used avail capacity Mounted on
sulky25:/AWG18 619788288 616803143 2800045 100% /a/sulky25/AWG18
sulky26:/AWG23 619788288 613374726 6013098 100% /a/sulky26/AWG23
sulky13:/AWG37 619788288 607730751 11310159 99% /a/sulky13/AWG37
They contain a mixture of files:
-----------------------------------------------------------------------------
Disk User Size What
-----------------------------------------------------------------------------
AWG18 vub-recoil 515948923 ISL/sx-080702: gen BB MC
data
skimmed data rfiles
moments -
miscellaneous
AWG23 vub-recoil 107038654 ISL/sx-080702: signal MC
cocktail MC
skimmed MC rfiles
hbook_gen_mar ???
ISL/VubAna-out: DELETE ???
moments 498613341
AWG37 vub-recoil 262869337 ISL/sx-080702: skimmed MC rfiles
ISL/sx-080702/newgenbb: gen BB MC
vub-combo 152832512
moments 191845657
-----------------------------------------------------------------------------
In vub-recoil space the following HBOOK files are around (why??):
----------------------------------------------------------------------
AWG23: ca 50GB in ISL/sx-080702/hbook_gen_mar/ and
ISL/sx-080702/newsig/output/outputdir/
AWG18: ca 12GB in ISL/sx-080702/newsignal/output/outputdir/ and
ISL/sx-080702/signalMC/FIXED/
----------------------------------------------------------------------
There is a minor problem in that many of the files and directories
belong to Alessio, and his account is gone. Will have to send a mail
to SCS.
We also occupy diskspace of group EC:
shire01>du -ks /nfs/farm/g/ec/u05/users/*
2126409 /nfs/farm/g/ec/u05/users/asarti
509917864 /nfs/farm/g/ec/u05/users/henning
vub-recoil does not use this diskspace, I think (I found no symlinks
to this disk and nothing in our chains). The story is different for
the moments-analysis, I guess.
-> There are hbook files to be deleted, and some logfiles are not
gzip'ed. We can free up >50 GB on relatively short notice, I guess.
2. CPU and Disk Usage for old production
========================================
I assume that we dump identical ntuples (root format) as before.
o ca. 25% of the generic BB allevents make it into the ntuples.
o Without tagbits, the jobs are CPU limited:
barb: 1.4sec CPU/allEvent
noma: 0.7sec CPU/allEvent
The wall clock time is only marginally (<5% .. 10%) higher. We'll
have to see how this is with a CM2 executable on OBJY.
o Running with tagbits was mostly relevant in the old (K0S bug
affected) MC. I cannot find tcl, log and rootfiles for this case.
Since the jobs are CPU limited I assume that the time gain is
proportional to the events skipped.
The skim fractions are 20% (somewhat less than what I see above,
but this may be due to updated(?) purity tables), see the full list
in http://www.slac.stanford.edu/BFROOT/www/Physics/Analysis/AWG/EHBDOC/skims/allskims.html
o Diskspace: 7.5 kB/tupleEvent
The above was generic BB MC. Now for data:
o CPU: 3.6sec CPU/filterEvent, 0.16sec CPU/allEvent,
o Disk: 4.5 kB/tupleEvent
3. Projection for new production
================================
Data: assume 150/fb skimmed CM2
-------------------------------
There are 1.5e7allEvents/fb-1 (see http://www.slac.stanford.edu/BFROOT/www/Computing/DataQuality/SkimmedData.html)
This implies for the semiExcl skim (with a rate of 4% on data)
-> 150 * 1.5e7 * 0.04 = 9.0e7 events
-> 9e7*4.5kB = 4.1e8kB = 410 GB
Assuming 1sec CPU/event, this is 25000 CPU hours. (This assumption may
be somewhat optimistic with CM2 ...) This is what a user can get from
the compute farm in about ONE month (if "run" is running 24/7).
generic BB MC: first assume 400/fb unskimmed SP5 OBJY
-----------------------------------------------------
This is ca. 420e6 allEvents. Assume 20% with semiExcl BRECO skim.
-> 0.2 * 420e6 = 8.4e7 events
-> 8.4e7*7.5kB = 630GB
Assuming again 1sec CPU/event and the availability of the tag bits,
this is 23000 CPU hours. If the "skim rate" is higher (because the tag
bits are based on older purity tables(?), the disk space could
increase by up to 25%, to 780GB).
If the tagbits are not available, the total time would be 125khours,
doable for a team in one month.
Now add 120/fb of skimmed CM2 SP6
---------------------------------
This is ca 24e6 events for the ntuple
-> 24e6 * 7.5kB = 180GB
The time to run on this is small, O(6000h).
|