The Downloading of SVT
L.Ristori
27-Jul-1995
1. Introduction
SVT contains a number of RAM's that need to be initialized.
Most of them will be downloaded at power-up and will never
change during normal running. In this note we make an
inventory of all these RAM's according to the present
status of the SVT design and estimate the time it will take
to initialize the whole system based on the amount of data
to be transfered and the expected available bandwidth.
2. Data Inventory
2.1 Hit Finder
Each Hit Finder reads all the R-Phi strips in a single wedge
(1/12 in Phi and 1/6 in z). There are about 2k such strips
in a Wedge. The pedestals of all the strips are stored in a
table. The size is 1 byte per strip.
Each hit Finder also has a Memory LookUp (MLU) to compute
the centroid of each cluster. It takes 5 bits each from
three consecutive strips as address an outputs 8 bits of
data (4 bit mantissa + quality flags). The size is 2**15 =
32k bytes.
2.2 Hit Buffer
The SS Map is addressed by 17 bits from the hit coordinate
and outputs a 16 bit pointer to the Hit List Memory. The
size is 2**17 x 16 bit = 256k bytes.
The AM Map is addressed by a 17 bit wide road number plus a
3 bit wide layer number and outputs a 16 bit pointer to the
Hit List Memory. The size is 2**(17+3) x 16 bits = 2M
bytes.
2.3 AM Sequencer
The SS Map is addressed by 17 bits from the hit coordinate
(same as in the Hit Buffer) and outputs a 15 bit SuperStrip
number (including a 3 bit layer field). The size is 2**17 x
15 bit = 256k bytes.
The sequencer is implemented as a finite state machine
driven by a RAM look-up. It has 16 inputs and 20 outputs (3
bytes). The size is 2**16 x 3 bytes = 192k bytes
2.4 AM Board
Each Associative Memory board houses 128 AM chips with 128
roads each. Each road is 6 layers and there is a 12 bit
SuperStrip address for each layer. The total size is then
128 x 128 x 6 x 2 bytes = 192k bytes.
2.5 Track Fitters
We do not know how long the algorithm will be but it has to
be very simple and a reasonable guess is that it will be
less than 1k.
If we assume that we will use for the fit 4 Si layers plus 2
parameters from the central tracker, we will have 3
constraints with 6 coefficients plus 2 cuts each and we
will need one set of 6 coefficients plus 1 offset for each
of the 3 parameters to be computed (phi, curvature, d). The
total set of geometrical constants is therefore 3x8 + 3x7 =
45 numbers. If we assume each number is 8 bytes we get 45x8
= 360 bytes. We will need one such set for each barrel
(x6), for each phi sector (x12), and for each possible
mixture of layers in adjacent barrels a track can traverse
(6 combinations x 5 inter barrel boundaries). So the total
size is 360 x 12 x (6 + 6x5) = 152k bytes.
3. Network Topology
In the attached figure we show the topology of the network
of CPU's that will control the operation of SVT. Each VME
crate has its own control CPU that is capable of accessing
all the modules in the crate via the backplane bus. All the
control CPU's are connected through a dedicated Ethernet
branch to a number of control workstations. A bridge
connects the SVT Ethernet branch to the rest of the world
and filters out all unwanted traffic.
4. Downloading time
All the data base needed for SVT is resident on mass storage
(disk) local to the control workstations. At initialization
time data are retrieved from disk by software running on
the control workstations and sent over Ethernet to the
control CPU's that proceed to dowload and initialize
individual modules.
In order to estimate the time it takes to initialize SVT, we
assumed that each RAM could be downloaded with a single I/O
transaction as seen by the control workstation operating
system. We also assumed that each transaction will require
a system overhead of 1 ms plus a time proportional to the
amount of data transfered and corresponding to 20% of the
total Ethernet bandwidth (20% of 10 Megabit/s). These
numbers are reasonable guesses based on past experience
with similar systems. We hope we will be able to measure
the real performance of this system in the near future and
to come up with more reliable estimates.
The table below shows the time estimates based on the
numbers given above broken down by module. The first column
(mult) is the total number of modules of that kind in SVT,
the second column is the module name, the third column
shows the data block to be downloaded, the fourth column
(kB each) shows the size, in kBytes, of each module, the
fifth column (kB x mult) shows the total size in all
modules, the sixth column (ms ea.) shows the time it takes
to download one module in milliseconds, the seventh column
(ms x mult) shows the time it takes to download all
modules. The last row shows the total amount of data (in
kBytes) to be downloaded and the total time it takes (in
milliseconds).
The total time to initialize is about 160 seconds. We believe
this is acceptable since it will be needed very seldom,
typically only as part of a system cold start at power on.
_________________________________________________________________
mult module data kB each|kB x mult| ms ea.| ms x mult
-----------------------------------------------------------------
72 Hit Finder Pedestals 2 144 9 648
72 MLU RAM 32 2,304 129 9,288
12 Hit Buffer SS Map 256 3,072 1,025 12,300
12 AM Map 2,000 24,000 8,001 96,012
12 AM Sequencer SS Map 256 3,072 1,025 12,300
12 Seq. program 192 2,304 769 9,228
24 AM Board Roads 192 4,608 769 18,456
12 Track Fitters Algorithm 1 12 5 60
12 Geom. const. 152 1,824 609 7,308
-----------------------------------------------------------------
TOTAL kB 41,340 ms 165,600
_________________________________________________________________