Date: Mon, 25 Sep 2000 17:01:30 -1000 (HST)
From: Tony Denault <denault@jeans.ifa.hawaii.edu>
To: John Rayner <rayner@jeans.ifa.hawaii.edu>
Cc: techgroup@jeans.ifa.hawaii.edu
Subject: Spex Time & Efficiency.



We obtain a spex request centered on taking as many cycles as
quickly as possible. I was asked to take a look at how to increase
efficiently and analysis factors related to overhead. CC's to the
techgroup since others are interested in instrument efficiency.

As a test case we took data under the following conditions:

   array size = 1024x1024
        itime = 5.0 seconds.
        coadd = 1
     Autosave = on
  Other parameters using the current spex defaults:
    slowcnt=12; ndr=32;

For my test, I took 10 cycles and averaged the time per cycles.

1. Test 1 - Use the normal spex default.
   The mean cycle time was 11.32 seconds.

2. Test 2 - Turn off TCS communications.
   The mean cycle time was 11.07 seconds.

3. Test 3 - Turn off DV display of images:
   The mean cycle time was 10.78 seconds.

4. Test 4 - Save to local disk, rather that stefan's network disk.
   The mean cycle time was 10.54 seconds.

5. Adjust NDR from 9 to 5, slowcnt from 12 to 5. (Note: the default
   NDR is 32, but at 5.0 itime only 9 NDR are done).
   The mean cycle time was  6.70 seconds.

6. Adjust NDR to 1, and slow counts to 1. (Fastest possible)
   The mean cycle time was 5.72 seconds.


At the fastest time the overhead was 0.72 seconds. Further probing
brakes down this overhead as:

   .100 - Array readout time.
   .141 - VME Transfers
   .073 - sort of pixel data into image.
   .340 - saving data to FIT file.
   .066 - misc overhead.

Or the fastest cycle time can be estimated by:

  cycle_time = itime + array_readout_time + 0.620 second fixed overhead.

These options can be added (but will add overhead to the cycles efficiency):

  a. Having TCS info in the FITS files adds 0.25 seconds.

  b. Having DV display images during the GO adds 0.30 seconds.
     (My test was done using an xserver on a hilo PC. I'm sure
      running on X on stefan would be much faster ).

  c. Saving to a network disk (/scrs1) rather that a local disk (/data) adds 0.24 seconds.

  d. The largest contribution is the NDR/Slow Count overhead. The defaults are
     SlowCount=12, NDR=(as much as possible)32.

     You must choose between read noise vs. efficiency.

     At 12 slowConnts the array_read_time is .355 seconds for a single NDR. If you boost
     NDR as high as possible, your array_read_overhead is going to be equal to your
     intergration time ( always > 50% efficiently).
     At 32 NDR it would take 12 seconds before itime > NDR_time. Or array_readout is
     equal to minimum_of( 12, itime ).


Tony

/--------------------------------------------------------------------------\
| Tony Denault                      |   Email: denault@irtf.ifa.hawaii.edu |
| NASA IRTF, Institute of Astronomy |                Phone: (808) 933-0893 |
| 640 North Aohoku Place            |                  Fax: (808) 933-0737 |
| Hilo, Hawaii 96720                |                                      |
\--------------------------------------------------------------------------/