This page is part of the ING manual
INS-DAS-23 Architecture for UltraDAS.
This section is copied unchanged from issue 1.1 of the ADD,
which was written before UltraDAS was constructed; hence the use of the
future tense and the general air of anxiety. The risks were contained successfully
in system s9, but they remain relevant should the system be revised in
any significant way.
The chosen architecture inherits many proven design-patterns, but it also
introduces new elements of design and is forced to reuse some patterns
that have given problems in the past. This section lists the major risks
that the finished system might not work satisfactorily. Risks that a working
system might be delivered late, and having cost too much, are not considered
Hard limits on readout speed
If the fibre-optic interfaces to the detector controllers do not have enough
bandwidth, the controllers must be detuned to less than the best rate that
the detectors can deliver. This problem is discussed in section 6 and we
have these findings:
There is no problem for all existing CCD cameras.
Cameras with only one channel can run up to 4.9 times as fast as the current
EEV42 CCDs before we run out of bandwidth.
A 12Kpixel-square mosaic of EEV42 CCDs cannot be run at full speed from
an Ultra 10 DAS-computer if standard camera-interfaces are used. This camera
can run at full speed if (a) each PCI-interface board runs two controllers
at full speed or (b) we use a DAS computer with more PCI slots.
A 16Kpixel-square mosaic of EEV42 CCDs needs both the dual PCI-interface
boards and more PCI slots to run at full speed.
Loss of pixels through DMA failure
If the buffer memory in the PCI interface overflows, pixels are lost. This
could happen if the transfer of pixels from the buffer across PCIbus to
SPARCstation memory does not keep up with the input of pixels from the
camera. It is hard to guarantee this real-time performance on Solaris;
doing so limits the form of the acquiring programme quite severely. UltraDAS
makes no special concessions to hard real-time input, but assumes that
there is enough buffer memory on the camera-interface board for an entire
readout; UltraDAS always waits for the buffer to empty between readouts.
There are two sizes of buffer memory in the interface boards. The standard
board has 16 Mwords which is enough to store one frame from an EEV42 or
one exposure from INGRID, but not enough to store an observation
from the INT WFC mosaic. The larger board, to be made specially for ING,
has 96 MWord of memory and the planned combinations of these boards will
store one readout of any planned camera.
Initially, both single-chip cameras and INT WFC will be run with the
standard interface. Hence, INT WFC could overflow the buffer either
randomly or systematically. A prototype should be made to check that this
does not happen.
Bottlenecks in the disk system
The throughput required of the data discs is discussed in section 6. To
exploit this throughput, the DAS server-programmes have to run efficiently.
Tests in Cambridge suggest that the I/O performance can fall off badly
for large images if the DAS computer is short of RAM. The system may slow
down by three or four times. More prototypes and tests should be carried
out to find how much memory is needed to prevent this.
The most critical test is to determine the speed at which data can be
written to a single partition on DASAN.
Preprocessing of images
The IR cameras require preprocessing of the data in memory before writing
it to disc. For UltraDAS, this means that the preprocessing is probably
going on in parallel with writing to disc. The implications of this on
memory usage and other system resources need to be checked. There is a
danger that the cycle time will get to be too long.
Both the hardware and the software for the PCI interface are produced by
SDSU. There is a danger that vital functions will be left out initially
and cannot be added later because ING does not have full knowledge of the
design or does not possess the source code. To avoid this, a protoype of
the server applications should be built to reveal the need for any obscure
Long communications delays between the clients and servers or between the
servers and controllers could spoil the efficiency of the system. Both
interfaces should be prototyped as soon as possible.