- Located at Wide Band
- At least 500 SPEC = Specint92 = VUPS (we will use the term SPEC hereafter)
- At least 100 SPEC per CPU
- FNAL supported farm/analysis system (AIX/HP-UX/IRIX/SunOS/OSF1), probably not IRIX
- Single SMP system, or Farm of workstations
- Receives all events from data logger via network
- Quick analysis of about 15% of all events, selected for richest physics.
- Data Tape output under 10% of raw data
- Oddpack and Dreampack alignment runs
- Could help with other online minitoring ( RECON, GHOST, LAZRUS et.al. )
In the 1990/91 E687 run, the ability to analyze small samples of data quickly in the counting
room played a major role in our ability to keep the quality of the data high.
We also derived some advantage from the "expressline" Silicon Graphics system in the Feynman
Center, used to run the occasional full data tape to theck charm signals, and used to
reconstruct some special calibration runs shortly after they were taken.
For the E831 data run we plan to have a much more powerful Express Analysis system in the
counting room.
Based on the projected budget and typical systems presently available, we can
expect at least 500 SPEC of CPU power.
The system should be configured with appropriate disk and tape,
for staging of data and logging of results.
We plan a data taking rate of about 20,000 events per spill, at about 5 KBytes per event.
This implies a sustained data rate of about 1.6 MBytes/second and 333 events/second.
The full analysis of an average E687 event requires about 20 SPEC-seconds.
Thus, if everything remained the same for E831, a 500 SPEC farm could analyze 7% of the data.
( Event sizes may increase due to the new Vertex and Straw detectors, and analysis time
could increase somewhat. )
While this fraction of the data corresponds roughly to a full analysis of E687, and would
provide good online charm signals and powerful diagnostics, there may be even better
stategies for use of the Express system. We would like to get a larger
fraction of the charm signal for monitoring, and for early access to Physics signals.
The full analysis of E831 events will be slower than for E687, due to the
additional detectors and improved efficiencies.
But full calorimeter reconstruction will
probably not be available, nor will it be necessary for most of the Express monitoring.
Streamlined calorimitry would probably bring the Express reconstruction down to half
the time of full E687 reconstruction, allowing 15% of the events to be analyzed.
We can make physics based event selections, through
Online
and
Offline
methods.
A combination of online and offline selections could be used.
If we are clever, we could see a large fraction of the final charm signal in some modes.
The Express system should write output tapes, mostly in DST format to keep
down the tape and data handling costs.
We will write no more than 1/10 as many tapes as the full data stream.
With all events available, a rapid partial analysis could select events for complete analysis.
Selections could include a combination of:
- Partial silicon tracking for vertex selection
- Et selection based on calorimeter data. A factor of three event selection keeps 80% charm.
- Electron and muon selections
Trigger bits from Level II may provide more Et, muon, and other information than we use to
select events for logging. These trigger bits could be used in Expess event selection.
We will discuss here the
analysis model,
data links,
farm configurations and
peripherals.
Some
reference systems will be described, based on vendor's catalogues.
These descriptions are not necessarily complete or correct, and other considerations may
lead us to alternate systems.
Let us assume, for reference, that we are working with a $150,000 total budget.
Based on Farm experience it is reasonable to plan on substantial peripheral
support, so that the available CPU can be used efficently.
Let us assume a budget of
- $ 120K - processors, memory, system disks and licenses,
- $ 30K - tape drives and data disks
Input data files will be stored on disk in the Express system.
The slowest individual CPU's we should consider are 100+ SPEC's,
comparable to the 300 SPEC power of a typical farmlet used in the E687 analysis.
We should not need formal farm managent systems like CVS, PVM
or SHIFT. Each input file can be analyzed by a single process.
We can afford to store several tapes of input and output data on disk,
decoupling Express processing from media handling.
Reading data into the Expresss system from tape would be expensive in terms of labor,
drive maintenance, and risk to the tapes.
Available network links permit the data
to be obtained directly from the Logger.
It appears that several good options are available.
We may need to add memory to the Logger to provide adequate decoupling of the
Express and Logging functions and to reduce Logger bus activity.
Some network options:
- Ethernet - a nominal sustained rate of 1.6 MBytes/second realistically requires
three dedicated ethernet links.
This solution does not scale well as data rates increase.
- Fast-ethernet - a single Fast-ethernet link should provides over 5 MBytes/second.
Fast-ethernet is relatively new, but seems to be widely available and inexpensive.
- FDDI - the same data rates as Fast-ethernet, again widely available.
FDDI interfaces are in some cases unreasonably expensive ($8 K for a $6K Indy),
and many adaptors impose an unacceptable load on the host.
- SCSI - Fast and Fast/Wide SCSI at 10 and 20 MB/sec is free and efficient.
The only barrier is the lack of software support.
Some vendors are starting to support SCSI clusters.
- BIT-3 link - Inexpensive ($3K/pair) bus adaptors can move data at lest 15 MB/sec
between almost any pair of available busses, with very small load on the host.
This is less flexible than general network solutions, but is a realistic and tested
fallback solution.
Multiple CPU's are required in order to obtained the required total computing power.
CPU's may be configured with Shared Memory on a shared high speed bus (SMP),
or may be purchased as standalone workstations in a Farm.
On an SMP system we will simply start one analysis task per CPU, and let the system
take care of scheduling and communication.
On a Farm system, a Host node will NFS serve the data disks.
We will run one analysis task on each Worker node.
A fast network hub connects to the Host via FDDI or Fast Ethernet,
connects to the Workers via regular ethernet,
and buffers the farm from the general E831 network.
We plan to produce only about 200 KBytes/second of output.
A single EXB-8505 will suffice for output logging.
A second drive is required for backups, and for a hot spare.
Three more drives, for a total of 5, will allow us to perform reanalysis during
beam down periods. The five drives should cost about $10K.
The highest data rates are on the disks holding the input data files.
To avoid head contention, we must try to analyze from one disk while filling another.
If analysis keeps well ahead of the data source, this requires two input disks.
It is quite likely that we will tune analysis to keep only slightly ahead of
the data. Therefore we need three input data disks.
Output data rates are low, but head contention can still prevent tape streaming.
Therefore we should plan on two output data disks.
Five data disks, at about $3K per 9 Gb, would cost about $15K
We should plan on at lest $5K for cables, chassis, etc for the peripherals.
For the remaining $120K, we list some reaonably cost-effective systems,
presently orderable from each of the Fermilab supported vendors.
We may not have a free choice of vendors, and the information listed
here may be incomplete.
Both small-node Farm and SMP systems are listed when available.
Published retail prices for standard configurations are used.
I have omitted Fermilab discounts, as well as software, maintenance, networking
and other infrastructure costs.
These factors may roughly cancel out, but the Devil is in the Details.
The numbers listed below should be used only to get a rough idea of how well we might do.
Systems are listed in inverse alphabetical order.
- Vendor - Model - $/node x nodes - nodes x (CPU/node x SPEC/CPU) = total SPEC
- SUN - SPARCstation 71MP . $25K x 5 . 5 x (2x125) = 1250 SPEC
- Sun's SPECfp figures are poor, 70% of the usual
- SPEC based on APOGEE compiler, K&A preprocessor
- Sun's fastest 2000E SMP system is 85 SPEC/CPU
- SGI - Challenge L . $120K .......... 1 x (4x130) = 520 SPEC
- SGI - Challenge S . $ 17K x 6 ...... 7 x (1x110) = 770 SPEC
- IBM - SP-2 ........ $140K .......... 1 x (2x100) = 200 SPEC
- additional CPU's cost $75K
- IBM - 59H ......... $ 42K x 3 ...... 3 x (1x130) = 390 SPEC
- HP. - HP9000 K400 . $ 65K x 2 ...... 2 x (4x135) = 1000 SPEC
- HP. - HP J210 ..... $ 55K x 2 ...... 2 x (2x170) = 680 SPEC
- DEC - Alphaserver 8200 . $130K ..... 1 x (2x341) = 682 SPEC
- additional CPU's cost $65K/2CPU ($94/SPEC) to total of 6 CPU
- DEC - Alphaserver 2100 . $60K+66K .. 1 x (3x277) = 831 SPEC
- additional CPU's cost $33K ($120/SPEC) to total of 4
- DEC - AlphaStation 250 4/266 $20K .. 6 x (1x200) = 1200 SPEC