IPAC 2MASS Working Group Meeting #87 Minutes
IPAC 2MASS Working Group Meeting #87 Minutes, 2/27/96
Attendees: R. Beck, R. Cutri, T. Evans, J. Fowler, L. Fullmer,
T. Jarrett, D. Kirkpatrick, G. Kopan, B. Light, H. McCallon
AGENDA
- Updated 2MAPPS Schedule
- User CPU Variation Report
- Ultrasparc Loaner
- Coadd Kernel Study
- WSDB SIS Definition
DISCUSSION
- Updated 2MAPPS Schedule -- The task of updating
the 2MAPPS schedule discussed at the previous meeting
has been completed, and copies of the new schedule were distributed to the
members present. Copies were put in the mailboxes of members not present.
- User CPU Variation Report -- G. Kopan
reported on a test he has been doing to study the variation of user CPU
time consumed on identical tasks. It appears that consistent access of
memory locations spanning over a megabyte thrashes the processor cache,
and once such thrashing begins, the user CPU, system CPU, and
throughput time vary significantly. If the memory spanned is just under
a megabyte, no thrashing is seen, and the times are all nearly constant
over repeated runs of the identical task. The test program and output
are appended to these minutes. The implications are that wherever
possible, consistent access of memory locations more than a megabyte
apart in memory should be avoided (note that this could take the form
of a window moving through memory with accesses occurring at the
beginning and end of the window and a width of over 1 MB). J. Fowler
will look into ways to reduce such access in DFLAT, where the
processing of 256x256 REAL*4 arrays involving the use of five arrays at
a time might be reorganized.
- Ultrasparc Loaner -- We have been asked by
the Tools Group how long we need the Ultrasparc loaner from Sun and
when we need it. It was not clear whether extra time implied extra cost
to IPAC. J. Fowler agreed to get these details. [After the meeting, R.
Scholey provided the following information: The loaner involves no cost
to IPAC, but Sun is reluctant to make it available for more than one
week; it is expected to arrive within a week or so.]
- Coadd Kernel Study -- T. Jarrett reported on
a study of the effects on galaxy processing that result from varying
the coadding kernel. Three approaches have been used: bilinear
interpolation, area overlap, and the "Weinberg kernel". The latter
involves smoothing the input before coadding, and the smoothing
function can be anything, but a Gaussian comparable in size to a
typical 2MASS point spread function was used in the tests. The area
overlap kernel produces the sharpest images (i.e., images appearing to
contain more power in high spatial frequencies). The Weinberg kernel
provided slightly more faint galaxies when the output was processed by
GALWORKS, but slightly more false extractions also resulted. The
differences between the three kernels were very small. A more complete
report is being prepared by T. Chester.
- WSDB SIS Definition -- R. Cutri discussed the
data to be included in the Working Survey Data Base (WSDB) needed to
support the final data products. He has put together a draft
specification that identifies parameters and formats. This will be
distributed, and ideas for adding, removing, or changing items are
solicited.
One aspect that became clear was that the size of the point source
record that BANDMERGE has to handle will probably become larger,
making it possible for memory allocation to be a problem. As a result,
J. Fowler requested that the MAP01 SIS be changed to include a
requirement for each line of ASCII text to be the same length so that
the corresponding files could be read as direct access files. This will
allow BANDMERGE to read only the parts of the files needed for band
merging and then to include the other information efficiently in the
output. T. Evans stated that this would not be a problem, although it
would require a blank space at the end of each line to correspond to
the final "|" delimiter in the header lines, but otherwise the only
change would be to pad out the comment lines to the same length as the
data lines.
Team members are requested to review the sample WSDB to be
distributed by R. Cutri and send responses to him.
Addendum: Code and Results From User-CPU Tests
c thrash the cache
parameter (nbuf=2**20)
byte s(nbuf),d(nbuf)
loop=2**24/nbuf
print *,loop,nbuf
do i=1,nbuf
s(i)=1
enddo
do j=1,loop
do i=1,nbuf
d(i) = s(i)
enddo
enddo
stop
end
4 4194304 8.88u 0.37s 0:09.36 98.8%
4 4194304 17.72u 0.46s 0:18.41 98.7%
4 4194304 17.74u 0.49s 0:18.41 99.0%
4 4194304 17.85u 0.38s 0:18.43 98.9%
4 4194304 17.53u 0.44s 0:18.24 98.5%
4 4194304 17.86u 0.46s 0:18.48 99.1%
4 4194304 17.82u 0.49s 0:18.56 98.6%
4 4194304 17.93u 0.40s 0:18.47 99.2%
4 4194304 17.88u 0.47s 0:18.64 98.4%
4 4194304 17.92u 0.46s 0:18.56 99.0%
8 2097152 5.46u 0.27s 0:05.79 98.9%
8 2097152 16.21u 0.20s 0:16.54 99.2%
8 2097152 16.24u 0.20s 0:16.55 99.3%
8 2097152 16.61u 0.26s 0:16.97 99.4%
8 2097152 16.79u 0.21s 0:17.13 99.2%
8 2097152 16.68u 0.25s 0:17.07 99.1%
8 2097152 16.68u 0.23s 0:17.06 99.1%
8 2097152 16.66u 0.26s 0:17.06 99.1%
8 2097152 16.75u 0.28s 0:17.13 99.4%
8 2097152 12.44u 0.21s 0:12.79 98.9%
16 1048576 19.74u 0.11s 0:19.94 99.5%
16 1048576 19.74u 0.16s 0:20.11 98.9%
16 1048576 19.71u 0.17s 0:20.08 99.0%
16 1048576 19.80u 0.14s 0:20.06 99.4%
16 1048576 19.74u 0.10s 0:19.99 99.2%
16 1048576 19.75u 0.11s 0:19.96 99.4%
16 1048576 19.82u 0.06s 0:19.98 99.4%
16 1048576 19.49u 0.41s 0:20.00 99.5%
16 1048576 19.84u 0.11s 0:20.07 99.4%
16 1048576 19.82u 0.15s 0:20.08 99.4%
32 524288 4.60u 0.08s 0:04.81 97.2%
32 524288 4.73u 0.12s 0:04.94 98.1%
32 524288 4.54u 0.15s 0:04.87 96.3%
32 524288 4.64u 0.11s 0:04.93 96.3%
32 524288 4.69u 0.09s 0:04.86 98.3%
32 524288 4.69u 0.09s 0:04.84 98.7%
32 524288 4.73u 0.09s 0:04.90 98.3%
32 524288 4.76u 0.07s 0:04.91 98.3%
32 524288 4.66u 0.10s 0:04.86 97.9%
32 524288 4.76u 0.12s 0:04.99 97.7%
64 262144 4.62u 0.05s 0:04.70 99.3%
64 262144 4.64u 0.08s 0:04.82 97.9%
64 262144 4.65u 0.04s 0:04.78 98.1%
64 262144 4.65u 0.07s 0:04.97 94.9%
64 262144 4.57u 0.10s 0:04.95 94.3%
64 262144 4.55u 0.09s 0:04.71 98.5%
64 262144 4.65u 0.05s 0:04.78 98.3%
64 262144 4.60u 0.08s 0:04.78 97.9%
64 262144 4.59u 0.08s 0:04.75 98.3%
64 262144 4.61u 0.08s 0:04.74 98.9%