LHCbComputing Model
Domenico Galli, Bologna
INFN CSN1
Roma, 31.1.2005
LHCb Computing Model. 2Domenico Galli
Premise: The Event Rates Current LHCb computing model and resource estimates
are based on the event rates at HLT output following “re-optimized trigger/DAQ/computing”:
Maximize physics output given available/expected computing resources.
They are summarized on the following table:
Data rate
EventsCalibratio
nPhysics
200 Hz Exclusive B candidates Tagging B (core)
600 HzHigh mass dimuon candidates
Tracking bJ/X (unbiased)
300 Hz D* candidates PIDCharm (mixing &
CPV)
900 HzInclusive b candidates (e.g. b)
Trigger B (data mining)
LHCb Computing Model. 3Domenico Galli
The LHCb Dataflow
On-line Farm
CERN Tier-1s
CERN Tier-1s
Tier-2s
reconstruction
pre-selectionanalysis
RAWmc data RAW data
rDST
DST+RAW TAG
calibration data
MC On-line Farm
Physics Analysis
Local Analysis
n-tuple User TAGUser DST
TAGSelected DST+RAW
Paper
CERN Tier-1s
Tier-3s
CERN
Scheduled job Chaotic job
LHCb Computing Model. 4Domenico Galli
The LHCb Dataflow (II)
LHCb Computing Model. 5Domenico Galli
Event Parameters
Event Size [kB]
current
2008
RAW 25 25rDST N/A 25TAG 1 1DST 100 75MC DST 500 400
CPU [kSi2k•s/evt]curren
t2008
Reconstruction 2.4 2.4
Pre-selection analysis
0.6 0.2
Analysis 0.3 0.3 rDST (reduced DST): only enough
reconstructed data to allow the physics pre-selections algorithms to be run.
reconstruction
pre-selectionanalysis
RAWmc data RAW data
rDST
DST+RAW TAG
calibration data
Efficiency assumed
%
Scheduled CPU usage
85
Chaotic CPU usage 60
Disk usage 70
MSS usage 100
LHCb Computing Model. 6Domenico Galli
5.5 MSi2k;
1800 CPU (assuming PASTA/2006-2007forecast);
40 TB disk.
The On-line Event Filter Farm
CERNcomputing
centre
HLTb-exclusive 200 Hz
di-muon 600 Hz
D* 300 Hz
b-inclusive 900 Hz
rDST (25 kB/evt)200 Hz
2 kHzRAW (25 kB/evt)
60 MB/s2x1010 evt/a
500 TB/a
2 streams
1 a = 107 s over 7-month period
200
600
300
900
b-exclusive
di-muon
D*
b-inclusive
LHCb Computing Model. 7Domenico Galli
Reconstruction Evaluate:
Track position and momentum.
Energy of electromagnetic and hadronic showers.
Particle identification (e, γ, π0, π/K, μ).
Make use of: Calibration and alignment constants (produced from
online monitoring and/or offline from a pre-processing of data associated with the sub-detector).
Detector conditions (a subset ofExperimental Control System database).
reconstruction24 kSi2k•s/evt
calibration dataRAW data25 kB/evt
rDST data25 kB/evt
LHCb Computing Model. 8Domenico Galli
Reconstruction (II) Required CPU for 1 pass of 1-year data set: 1.5 MSi2k•a. Performed twice in a year.
pass
1
during data taking, over a 7-month period
b-exclusive events: real-time, by the Event Filter Farm.
b-inclusive, di-muon, D* events: quasi real-time, (maximum delay of few days) by the Tier-1s.
CPU power required for each Tier-1: 0.39 MSi2k.
2
re-processing, during winter shut-down, over a 2-month period
42% by Event Filter Farm (5.5 MSi2k•2 months).
52% by Tier-1s and CERN.
CPU power required for each Tier-1: 0.74 MSi2k.
LHCb Computing Model. 9Domenico Galli
Reconstruction (III) 500 TB/a input RAW.
Stored on MSS in 2 copies: one at CERN, the other divided among Tier-1s:
500 TB/a @ CERN;
500/6 = 83 TB/a @ each Tier-1;
500 TB/a output rDST per pass. 1000 TB/a stored on MSS in 1 copy divided among CERN
and Tier1s: 1000/7 = 150 TB/a @ each CERN + Tier-1
LHCb Computing Model. 10Domenico Galli
Pre-selection Analysis (aka Stripping) Evaluate:
4-momentum of measured particletracks;
Primary and secondary vertices; Candidates for composite particles; 4-momentum of composite particles.
Apply: Cuts based on a specific
pre-selection algorithm foreach of the ~40 physicschannels.
At least 4 output data streamsforeseen during the first datataking (b-exclusive,b-inclusive, di-muonand D*).
rDST25 kB/evt
RAW25 kB/evt
TAG
b-inclusiveDST+RAW100 kB/evt
b-exclusiveDST+RAW100 kB/evt
D*
rDST+RAW50 kB/evt
di-muonrDST+RAW
50 kB/evt
pre-selectionanalysis
0.2 kSi2k•s/evt
Outputstream
Input fractio
n
Reduction
factorb-exclusive
0.1 10
b-inclusive 0.45 100di-muon 0.3 5D* 0.15 5
LHCb Computing Model. 11Domenico Galli
Pre-selection Analysis (II) Pre-selection cuts are looser with respect to final
analysis and include sidebands to extract background properties.
The event that pass the selection criteria will be fully reconstructed (full DST, 75 kB/evt).
An Event Tag Collection is created for faster reference to selectedevents; it contains:
a brief summary of eachevent’s characteristics;
the results of the pre-selectionalgorithms;
a reference to theactual DST record.
rDST25 kB/evt
RAW25 kB/evt
TAG
b-inclusiveDST+RAW100 kB/evt
b-exclusiveDST+RAW100 kB/evt
D*
rDST+RAW50 kB/evt
di-muonrDST+RAW
50 kB/evt
pre-selectionanalysis
0.2 kSi2k•s/evt
LHCb Computing Model. 12Domenico Galli
Pre-selection Analysis (III) Required CPU for 1 pass of 1-year data set: 0.29
MSi2k•a. Performed 4 times in a year.pass
1
during data taking, over a 7-month periodQuasi real-time (maximum delay of few days) by CERN + Tier-1s.CPU power required for each Tier-1/CERN: 0.08 MSi2k.
2after data taking, over a 1-month periodCPU power required for each Tier-1/CERN: 0.59 MSi2k.
3
after re-processing, during winter shut-down, over a 2-month periodCPU power provided by Event Filter Farm: 42% = 0.86 MSi2k.CPU power required for each Tier-1/CERN: 0.17 MSi2k.
4before next year data taking, over a 1-month periodCPU power required for each Tier-1/CERN: 4.1/7 = 0.59 MSi2k.
LHCb Computing Model. 13Domenico Galli
Pre-selection Analysis (IV) Input: 2 pass x 500 TB/a rDST.
Output: 4 pass x (119 + 20 = 139) TB/a DST+TAG.
Stored on MSS in 2 copies: one at CERN, the other divided among Tier-1s:
4x139 = 556 TB/a @ CERN;
556/6 = 93 TB/a @ each Tier-1;
Stored on disk in 7 copies: one at CERN, one for each Tier-1. Older version removed (2 version kept):
2x139 TB/a @ CERN;
2x139 TB/a @ each Tier-1;
LHCb Computing Model. 14Domenico Galli
Simulation Simulation studies are usually performed in order to:
measure the performance of the detector and of the event selection as a function of the regions of phase space;
estimate the efficiency of the full reconstruction and analysis of the B decay channel.
Due to the large background rejection, a full simulation of background events is unfeasible. Moreover it is better to rely on real data (mass sidebands) than on MC samples.
Simulation strategy: concentrate the simulation on what we consider as main-stream signals, in particular B decays and b-inclusive events.
Statistics must be sufficient such that the total error is not dominated by MC statistical error.
LHCb Computing Model. 15Domenico Galli
Simulation (II) 2•109 signal events;
2•109 b-inclusive events;
10% of these events will pass the trigger simulation and will be reconstructed and stored on MSS.
6.5 MSi2k•a required (dominates CPU needs for LHCb).
MC DST size (including “thruth” information and relationships) is ~400 kB/evt. TAG size is ~1 kB/evt.
MSS storage: 160 TB/a.
LHCb Computing Model. 16Domenico Galli
Analysis Analysis starts from stripped DST.
Output of stripping is self-contained, i.e. no need to navigate between files.
Further reduces the sample (typically by a factor of 5) to focus on one particular analysis channel.
Produce an n-tuple object or aprivate stripped DST, used bya single physicist or a small groupof collaborators.
Typical analysis jobs run on a~106 event sample.
Some analysis jobs will run ona larger ~107 event sample.
Physics Analysis0.3 kSi2k•s/evt
Local Analysis
n-tuple User TAGUser DST
TAGSelected DST+RAW
Paper
LHCb Computing Model. 17Domenico Galli
Analysis (II)
Estimate of analysis requirements, excluding efficiencies.
N. of physicist performing analysis 140 (25%)
N. of analysis jobs per physicist per week 4
Fraction of jobs analyzing 106 events 80%
Fraction of jobs analyzing 107 events 20%
Event size reduction factor after analysis 5
Number of active n-tuples 5
2008 CPU needs [MSi2k•a] 0.80
2008 Disk storage [TB] 200
LHCb Computing Model. 18Domenico Galli
Analysis (III) CPU need in 2008 (including 60% efficiency):
1.3 MSi2k•a. Due to better access to the RAW data, past copies of
stripped DST and the availability of MC data, we foresee CERN servicing a larger fraction of the analysis:
CERN: 25%; Tier-1: 75% (12.5% each one).
CPU power required in 2008 for CERN:1.3 * 0.25 = 0.32 MSi2k•a.
CPU power required in 2008 for each Tier-1:1.3 * 0.75/6 = 0.16 MSi2k•a.
CPU need for analysis will grow linearly with the available data in the early years of data taking (e.g. 3.9 MSi2k•a in 2010).
Disk storage need in 2008: ~200 TB. (will grow linearly with the available data in the early years of the experiment, e.g. ~600 TB in 2010).
LHCb Computing Model. 19Domenico Galli
Data Location (MSS) Tier-1s:
INFN-CNAF (Bologna, Italy)
FZK (Karlsruhe, Germany)
IN2P3 (Lyon, France)
NIKHEF (Amsterdam, Netherlands)
PIC (Barcelona, Spain)
RAL (UK)
DST x 2
rDST
Tier1’s
RAW x 2
CERN
MC x 2
LHCb Computing Model. 20Domenico Galli
2008 Assumed first year of full data taking:
107 seconds @ 2 x 1032cm-2s-1. Extended over 7 month (April-October) These are “stable running conditions”.
Data sample:
b-exclusive
dimuon D*b-
inclusiveTotal
Trigger Rate [Hz] 200 600 300 900 2000
Events [x109] 2 6 3 9 20
LHCb Computing Model. 21Domenico Galli
CPU Requirements in 2008
MSi2k•a CERN 6 Tier- 1s Tier- 1 14 Tier- 2s Tier- 2 Total
Stripping 0.17 1.03 0.17 0.00 0.00 1.20Recons. 0.40 2.42 0.40 0.00 0.00 2.83Monte Carlo 0.00 0.00 0.00 7.65 0.55 7.65Analysis 0.32 0.97 0.16 0.00 0.00 1.29Total 0.90 4.42 0.73 7.65 0.55 12.97
Online FARM resources not presented here. CPU efficiencies:
Production: 85%. Analysis: 60%.
LHCb Computing Model. 22Domenico Galli
CPU Requirements in 2008 (II)
0.00
5000.00
10000.00
15000.00
20000.00
25000.00
kSi2
k
Tier2's Tier1's CERN
LHCb Computing Model. 23Domenico Galli
CPU Requirements in 2008 (III)
0
5000
10000
15000
20000
25000
kSi2
k
Monte Carlo Analysis Reconstruction Stripping
LHCb Computing Model. 24Domenico Galli
Permanent Storage (MSS) in 2008
40%
60%
CERN
Tier1's
29%
33%
9%29%
RAWrDSTData DSTMC DST
TB CERN 6 Tier- 1s Tier- 1 Total
RAW 500 500 83 1000rDST 143 857 143 1000Data DST 556 556 93 1112MC DST 160 160 27 321Total 1359 2074 346 3433
LHCb Computing Model. 25Domenico Galli
Fast Storage (Disk) in 2008
74%
1%25%
CERNTier1'sTier2's
TB CERN 6 Tier- 1s Tier- 1 14 Tier- 2s Tier- 2 Total
RAW 136 0 0 0 0.0 136rDST 136 0 0 0 0.0 136Data DST 256 1534 256 0 0.0 1790MC DST 229 687 115 23 1.6 939Analysis 70 210 35 0 0.0 280Total 826 2432 405 23 1.6 3281
4% 4%
54%
29%
9% RAWrDSTData DSTMC DSTAnalysis
LHCb Computing Model. 26Domenico Galli
Network Bandwith
Peak bandwidth need exceed the average by a factor of 2.
[MB/s] CERN Tier-1 Tier-2
Average 76 143 20
Peak 165 276 20
LHCb Computing Model. 27Domenico Galli
Network Bandwith (II)
0.00
50.00
100.00
150.00
200.00
250.00
300.00
350.00
400.00
450.00
500.00
MB/s
CERN→ Tier1's→ Tier2's→
LHCb Computing Model. 28Domenico Galli
Network Bandwith (III)
(MB/s) Months Tier2's→ Tier1's→ CERN→J an-Mar 20 0 0Apr-Oct 20 39 34Nov 20 276 46Dec 20 128 165J an 20 128 165Feb 20 276 46Mar 20 0 0Apr-Oct 20 39 34Nov 20 276 46Dec 20 266 188J an 20 266 188Feb 20 276 46Mar 20 0 0Apr-Oct 20 78 40Nov 20 276 46Dec 20 266 188
2008
2010
2009
LHCb Computing Model. 29Domenico Galli
CPU growth
[MSi2k•a] 2006 2007 2008 2009 2010
CERN T0 + T1 0.27 0.54 0.90 1.25 1.88Tier1s 1.33 2.65 4.42 5.55 8.35Tier2s 2.29 4.59 7.65 7.65 7.65Total 3.89 7.78 12.97 14.45 17.87
0.00
5.00
10.00
15.00
20.00
MS
I2k
a
2006 2007 2008 2009 2010
Year
CPU need profile
CERN T0 + T1
Tier1s
Tier2s
LHCb Computing Model. 30Domenico Galli
Permanent Storage (MSS) growth
[TB] 2006 2007 2008 2009 2010
CERN T0 + T1 408 816 1359 2858 4566Tier1s 622 1244 2074 4286 7066Tier2sTotal 1030 2060 3433 7144 11632
0
2000
4000
6000
8000
10000
12000
TB
2006 2007 2008 2009 2010
Year
MSS need profile
CERN T0 + T1
Tier1s
Tier2s
LHCb Computing Model. 31Domenico Galli
Fast Storage (Disk) growth
[TB] 2006 2007 2008 2009 2010
CERN T0 + T1 248 496 826 1095 1363Tier1s 730 1459 2432 2897 3363Tier2s 7 14 23 23 23Total 984 1969 3281 4015 4749
0
1000
2000
3000
4000
5000
TB
2006 2007 2008 2009 2010
Year
Disk need profile
CERN T0 + T1
Tier1s
Tier2s
LHCb Computing Model. 32Domenico Galli
Cost Comparison200Hz 2000Hz
Hoffman2007
Now2008
CERNCPU [MSI2k] 2.0 0.9Disk [PB] 0.3 0.8Tape [PB] 1.2 1.4
Tier-1’s
CPU [MSI2k] 8.3 4.4Disk [PB] 1.6 2.4Tape [PB] 0.75 2.1
Tier-2’sCPU [MSI2k] - 7.6Disk [PB] - 0.02
Relative Cost 1.0 0.8
“Now” estimate based on CERN financing model (no internal LAN estimates though)
delay in purchasing, PASTA III report, …
Re-optimization Cost Comparison
LHCb Computing Model. 33Domenico Galli
Tier-2 in Italy In LHCb computing model Monte Carlo
production is performed at Tier-2s.
LHCb-Italy has currently no priority on Tier-2 resources.
We see the following options: Reserve some Tier-1 resources to perform
Monte Carlo production as well. Build-up LHCb Tier-2(s). Add resources for LHCb to existing Italian Tier-
2s.
LHCb Computing Model. 34Domenico Galli
Tier-1 In LHCb computing model Tier-1s are the
primary user analysis facility.
We need in Tier-1 fast random disk access.
We are investigating the solution of parallel file systems together with SAN technology as a mean to achieve the required I/O needs.
LHCb Computing Model. 35Domenico Galli
Testbed for Parallel File Systems @ CNAF
14 File Servers
40 GB IDE disk
GPFSPVF
SLustre
Gigabit switch
4 Gb trunke
d uplinks
Rack of 36
Clients
Network Boot
Server for the 14 File
Servers
Gigabit switch
LHCb Computing Model. 36Domenico Galli
Parallel File Systems: Write Throughput Comparison
Write aggregate throughput
0
50
100
150
200
250
1 5 10 20 30
Number of clients
MB
/s
PVFS-1 POSIX
PVFS-2 POSIX
GPFS POSIX
Lustre POSIX
PVFS-1 NATIVE
PVFS-2 NATIVE
LHCb Computing Model. 37Domenico Galli
Parallel File Systems: Read ThroughputComparison
Read aggregate throughput
0
50
100
150
200
250
300
350
400
1 5 10 20 30
Number of clients
MB
/s
PVFS-1 POSIX
PVFS-2 POSIX
GPFS POSIX
Lustre POSIX
PVFS-1 NATIVE
PVFS-2 NATIVE
Back-up
LHCb Computing Model. 39Domenico Galli
Trigger/DAQ/Computing Re-optimization In the original model only b-exclusive decays were
collected (200 Hz). The idea was to understand the properties of the
background by the simulation of large samples of background events.
In the meanwhile, also having in mind the Tevatron experience, we realized that whenever and as much as possible we need to extract information from real data itself.
E.g. study the background from the sideband of the mass spectrum
Collect unbiased samples of b events E.g. trigger on the semileptonic decay of the other B.
The net effect is the reduction of the CPU need for simulation but the increase in the storage need with no overall increase in the cost.
LHCb Computing Model. 40Domenico Galli
Dimuon Events Simple and robust trigger:
L1: ~1.8 kHz of high-mass dimuon trigger without IP cuts
HLT (with offline tracking and muID): ~600 Hz of dimuon candidates with high-mass (J/ or above)
Clean and abundant signals: J/, (1S), … Z mass peaks
Unique opportunity to understand the tracking (pin down systematics) Mass and momentum (B field calibration):
use dimuons from resonances of known masses
IP, decay length, proper time resolution: use J/ dimuons, which have common origin
Check of trigger biases: flat acceptance (vs proper time) for all BJ/X channels could be used as a handle to understand acceptance vs proper time
for other channels where IP cuts are applied
mass within 500 MeV of J/ or B mass, or above B mass
Huge statistics enables study as
a function of many parameters: geometry, kinematics (phase space)
LHCb Computing Model. 41Domenico Galli
Loose offline selection, after L0 and L1-dimuon without IP cut (no HLT yet)
~130 Hz of signal J/, dominated by prompt production
O(109) signal J/ per year = O(103) CDF’s statistics
Possible conceptual use for calibration of proper-time resolution (to be studied):
Bs Dsh CP/mixing fit sensitive to ~5% variations in the global scale factor on the proper-time resolution would help to know it to ~1% O(105) J/ needed for such precision.
To check event-by-event errors, extract scale factors in “phase space” cells can envisage up to 104 cells (e.g. 10 bins in 4 variables).
J/ signal
LHCb Computing Model. 42Domenico Galli
D* events Dedicated selection can collect abundant and
clean D* D0(K) peak without PID requirements
Such events can be used for PID (K and ) calibration + additional constraint for mass scale, etc …
Large statistics again allows study in bins of phase space
LHCb Computing Model. 43Domenico Galli
b events
Straightforward trigger at all levels: Require muon with minimum pT and impact parameter
significance (IPS)
Rely only on one track (robustness !)
No bias on other b-hadron Handle to study and understand our other highly-
biasing B selections
Example: Set pT threshold at 3 GeV/c and IPS threshold at 3
900 Hz output rate, including 550 Hz of events containing true b decay
LHCb Computing Model. 44Domenico Galli
Estimate of Reconstruction CPU & MSS
Off-line computing requirements for the reconstruction
b-exclusive
b-inclusive
di-muon
D* Total
Input fraction 0.1 0.45 0.3 0.15 1.00
Number of events
2.0•109 9.0•109 6.0•109 3.0•109 2.0•1010
CPU [MSi2k•a] 0.15 0.68 0.45 0.23 1.52
Storage requirement per reconstruction pass [TB]
50 225 150 75 500
LHCb Computing Model. 45Domenico Galli
Estimate of Reconstruction CPU & MSS (II) Required CPU for 1 pass of 1-year data set: 1.52
MSi2k•a. Performed twice in a year. 1st pass (during data taking, over a 7-month period):
CPU power required (assuming 85% CPU usage efficiency):(1.52-0.15)*12/7*100/85 = 2.8 MSi2k.
CPU power required for each Tier-1: 2.8/7 = 0.39 MSi2k. 2nd pass (re-processing, during winter shut-down,
over a 2-month period): CPU power required (assuming 85% CPU usage efficiency):
1.52*12/2*100/85 = 10.7 MSi2k. CPU power provided by Event Filter Farm: 5.5 MSi2k. CPU power to be shared between CERN and Tier-1s: 5.2 MSi2k. CPU power required for each Tier-1: 5.2/7 = 0.74 MSi2k.
LHCb Computing Model. 46Domenico Galli
Estimate of Pre-selection Analysis CPU & Storage
Reduction factors and computing requirements of the stripping stage
b-exclusive
b-inclusive
di-muon
D* Total
Input fraction 0.1 0.45 0.3 0.15 1.00
Reduction factor
10 100 5 5 9.57
Event yield per stripping
2.0•108 9.0•107 1.2•109 6.0•108 2.09•109
CPU [MSi2k•a] 0.03 0.06 0.13 0.06 0.29
Storage requirement per stripping [TB]
20 9 60 30 119
TAG [TB] 2 9 6 3 20
LHCb Computing Model. 47Domenico Galli
Estimate of Pre-selection Analysis CPU & Storage (II)
Required CPU for 1 pass of 1-year data set: 0.29 MSi2k•a.
Performed 4 times in a year.
1st pass (during data taking, over a 7-month period): CPU power required (assuming 85% CPU usage efficiency):
0.29*12/7*100/85 = 0.58 MSi2k.
CPU power required for each Tier-1/CERN: 0.58/7 = 0.08 MSi2k.
2nd pass (after data taking, over a 1-month period): CPU power required (assuming 85% CPU usage efficiency):
0.29*12/1*100/85 = 2.1 MSi2k.
CPU power required for each Tier-1/CERN: 4.1/7 = 0.59 MSi2k.
LHCb Computing Model. 48Domenico Galli
Estimate of Pre-selection Analysis CPU & Storage (III) 3rd pass (after re-processing, during winter shut-down,
over a 2-month period): CPU power required (assuming 85% CPU usage efficiency):
0.29*12/2*100/85 = 2.05 MSi2k. CPU power provided by Event Filter Farm: 42% = 0.86 MSi2k. CPU power to be shared between CERN and Tier-1s: 1.19
MSi2k. CPU power required for each Tier-1/CERN:
1.19/7 = 0.17 MSi2k.
4th pass (before next year data taking, over a 1-month period):
CPU power required (assuming 85% CPU usage efficiency):0.29*12/1*100/85 = 2.1 MSi2k.
CPU power required for each Tier-1/CERN: 4.1/7 = 0.59 MSi2k.
LHCb Computing Model. 49Domenico Galli
Estimate of Simulation CPU & MSS
Application Nos. of events
CPU time/evt [kSi2k•s]
Total CPU [kSi2k•a]
Signal Gauss 2109 50 3171
Boole 2109 1 63
Brunel 2108 2.4 15
Inclusive Gauss 2109 50 3171
Boole 2109 1 63
Brunel 2108 2.4 15
Total 6499
LHCb Computing Model. 50Domenico Galli
Estimate of Simulation CPU & MSS (II)
Output Nos. of events
Storage/evt [kB]
Total Storage
[TB]
Signal DST 2108 400 80.0
TAG 2108 1 0.2
Inclusive DST 2108 400 80.0
TAG 2108 1 0.2
Total 160.4
LHCb Computing Model. 51Domenico Galli
Estimate of Analysis CPU 140 physicists
4 jobs physicist-1 week-1
52 weeks/a
2.8x106 events/job
0.3 kSi2k•s/evt
efficiency = 0.6
jobs/a = 140 x 4 x 52 = 30000
CPU required in 2008:30000 jobs x 2.8x106 evt/jobs x 0.3 kSi2k•s/evt == 2.5x1010 kSi2k•s = 2.5x1010 kSi2k•3x10-8a = 0.80 MSi2k•a
CPU required in 2008 (including 60% efficiency):0.80 / 0.6 = 1.3 MSi2k•a