Download presentation
Presentation is loading. Please wait.
Published byMargery Fields Modified over 8 years ago
1
1 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 An Update to the CDF Offline Plan and FY2006 Budget ● Outline: – CDF computing model – Budget for FY'06 and projections for FY'07-08 P.Murat, FNAL
2
2 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CDF Computing Infrastructure ● CPU: Interactive and Batch Analysis Facilities ● Storage Systems ● Databases ● Networking Robotic Tape Storage System Reconstruction Farm Batch Analysis Farms FermiGRID GRID Disk Cache Analysis Disk Servers Central interactve login pool Databases Data Catalog
3
3 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Model of the CDF Computing: the inputs ● performance of the Tevatron – Luminosity profile: use design projections by the accelerator division (8.5 fb - 1 by 2009) – operations: ● 100 hours/week, 40 weeks/year, incorporate long shutdowns during the previous years ● performance of the CDF detector: – Peak logging rate: 20 MBytes/sec (2002-2004) –> 60 MBytes/sec (2007- 2008) – data logger upgrade in 2005-2006 – Overall operational efficiency: 50% (2002) -> 90% (2009)
4
4 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Model of the CDF Computing ● RAW event size: – reduced by a factor of 2 over the years – dependence on the instantaneous luminosity (new for this year) ● Profile of the instantaneous luminosity – = 0.4 L(peak) ● Dynamically prescaled triggers to utilize DAQ bandwidth more efficiently – average data logging rate: 60% of the peak ● 7 output data streams written by the data logger – 10% overlap between the 7 raw data streams – For 2 streams out of 7 do not drop results of the L3 reconstruction
5
5 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Model of the CDF Computing ● Production output : total volume 15% larger than the input volume ● MC data volume: 1.3 x reconstructed data volume MC event size = 1.3 x reconstructed event size (measured) N(MC events) = N(reconstructed events) ● Ntuple event size: 30% of the data/MC event (accounts for more than 1 commonly used ntuple format) – All the reconstructed and MC data ntupled
6
6 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 ● Cross-checks: 2 models using different parametrizations and simplifications: – Data rates (in MB/sec) out of the data logger (I) – event rates out of the trigger system (II) ● Overall agreement between the models within 10%, use (I) as more conservative asymptotically ● Primary source of uncertainty: expected volume of the RAW data Data volume estimates
7
7 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CPU Requirements: Offline Reconstruction ● 1-pass scheme: process the data 1.3 times (0.3 – calibrations) – output of the calibration processing is not stored – = 3.8 Ghz*sec – New in the model: event processing time dependence on the L inst – Assume luminosity profile doesn't change: = 0.4 L_inst (peak) – CPU needs well predicted – unplanned reprocessings: ~ 10% (based on operational 4 cycles)
8
8 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CPU Requirements: Off-site Computing ● Various types of jobs ● Biggest single consumer'2005: MC Production – ~600M events – ~ 30 Ghz*sec ● Individual user MC ● Analysis jobs, not requiring large amount of input – Example: top mass fitting ● Conclusion: – CDF is using offsite CPU resources efficiently SDSC CAF FY'2005: virtual machine allocation SDSC CAF in FY'2005: waiting jobs
9
9 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 ● Various kinds of jobs, the largest single one – making ntuples (~50% of the total) – Ntuple making executables run physics analysis algorithms. They are not fast ● High-Pt analyses: ~ 3GHz*sec – B-analyses: ~ 10 GHz*sec ● event size : ~ 20% of the reconstructed event ● 2005: ntuples are made at Fermilab ● Efficiency of using onsite CAFs ~95% CPU requirements: Fermilab CAF's
10
10 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 ● 2005: observe changes in CPU usage pattern: – Production: new version of the reconstruction algorithms (v6), process only 2005 data – generate run-dependent MC only for 2005 data, do not regenerate MC for the old runs – produce ntuples only for 2005 data ● The total dataset: – 2 parts (2002-2004 data and 2005 data), 50:50 approximately – For each part MC and the data processed with the same algorithms ● Moving from a mode where CPU needs scale with the total dataset size to a different mode where the component proportional to the added dataset size becomes significant CPU(2006) = A*D(total) + B*D(2006) ● Analyzed 2 extreme cases: A=0 and B=0 Model of the CPU Requirements
11
11 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CPU Requirements: “incremental” model ● CPU needs scale as an increment to the dataset size ● Large deviations from Moore's scaling (~x1.6 /year), use x1.3/year ● Used 2004-2005 data to tune the model of the CDF computing needs ● Plan to have ~50% CPU resources off-site starting from 2007
12
12 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CPU Requirements: “proportional” model ● CPU needs scale as the total dataset size ● Plan to have ~50% CPU resources off-site starting from 2007 ● Budget “proportional” model
13
13 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CPU requirements: ratio between the analysis and MC jobs ● Under this assumption: analysis needs : MC needs ~ 1:1 ● Primary reconstruction adds a relatively small fraction ● The most simplistic model: Total needs Production needs Analysis needs
14
14 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 New data on Analysis Farm Usage ● DCAF monitoring: offsite CPU facilities are extensively used for analysis ● FNAL: 20% MC vs 40% analysis, offsite: 60% MC vs 40% analysis ● Relative CPU consumption: Analysis:MC ~ 2:1 CAFs@FNALCAFs@FNAL: 3.4 THzAll DCAF's: 2.4 THz
15
15 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Disk Requirements ● Requirements model: total disk volume proportional to the dataset size ● Plan 2006 purchase to meet 2007 needs, retire 4-year old fileservers ● Price model: $20K per 14 TB fileserver in 2005, Moore's scaling (x2 in 18 month) ● 2 components: disk cache (backed up by the tape) and analysis disk (static) – Optimization strategies may be different ● Taking steps towards optimization of the disk usage: – Reorganizing project disk into the common pool – Will work with D0 and CD to develop disk cache simulation
16
16 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Tape Drive Requirements ● Recent decision: acquisition of a new tape robot in FY'2006, arrives in March ● LT0III drives, 60 MB/sec (assume $16K/drive+mover), in 2006 use 0.2 TB tape ● Expect that one-time doubling the tape I/O bandwidth will cover the needs of the experiment through 2009
17
17 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Interactive systems, DB, miscelaneous ● Interactive systems: new code servers, interactive Login Pool upgrade ● Databases: Frontier servers bought, just maintenance ● Miscellaneous: not predicted hardware procurements Central DB FronTier Squid FronTier Squid FronTier Squid FronTier Squid FronTier Squid FronTier Squid FronTier Squid USER JOBS SAM ● DB infrastructure: – Central server (ORACLE) – Replicas (FronTier squids)
18
18 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 CDF Networking'2005
19
19 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Networking Requirements ● FY'06 networking upgrades: – FCC/GCC : 3 rd 6509 switch + Starlink upgrade – CDF B0/trailers : switches and fabric upgrades
20
20 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Computing Requirements Summary ● “proportional” model used ● In 2007-2008 ratio about 50% of the total CPU located offsite
21
21 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Tapes and Operating Budget ● Capacity of 2 CDF STK silos - 10900 slots ● Budget STK9940B tapes ($75) for 2006, LTOIII tapes ($50) for 2007-2008 ● new LTOIII robot: $50/slot, 5 LTOIII drives budgeted for 2006 ● Contingency 15% included ● Recycling and compactification of the existing tapes in progress
22
22 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Equipment Summary ● the most expensive component is CPU (“proportional” model budgeted) ● CPU acquisition strategy assumes 50% of the total computing located offsite
23
23 P. Murat, Mini-review of the CDF Computing Plan 2006, 2005/10/18 Summary ● Improved, more detailed, model of the CDF computing ● FY'06 CDF computing budget request : – Equipment: $1,66M – Operating and tapes: $0.21M ● 'End game' scenario : 50% of the total CPU resources offsite
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.