Here are the results of some CPU estimates:

GEANT     180.00 +/- 50 ms/event 
ANALYSIS  150.00 +/- 50 ms/event  

GEANT was run on a CELERON 530. The analysis was run on mitch and
a PII 450 with similar results on both machines.

Assuming a P III 1.0 GHZ dual pentium would cut this time by a 
factor of 4 (2 because its a faster machine and 2 for the 2 CPU's):

GEANT     45.00 ms/event 
ANALYSIS  40.00 ms/event  

For 10^8 events:

GEANT	  52 days
ANALYSIS  46 days

I would think it is reasonable to consider between 5-10 such machines
in the coming year. I would tend to think, however, that it will
be about 1 year before we understand the detector/analysis well 
enough so as to need the full generation and analysis of 10^8 events.
Most of the calibrations will not require the analysis of this many
events. So we might want to delay the purchase of the machines
until we need them.

For the full analysis (several sets of 10^9 events) we will need at 
least 20 machines to make the analysis time reasonable, but hopefully by 
then CPU's would be few factors faster.


Uncertainties:
--------------

GEANT:

1. Rob has been working on changing the way the digitization is done
to make the Monte Carlo run faster, but he is still trying to fix a bug,
so we're not sure how much of an impact this will make (perhaps as much 
as 40 ms/event).

2. GEANT has not been optimized for CPU time. This remains a project that
needs to be done, and might be a good summer student project working closely
with Peter Gumplinger. It's not clear how much CPU we will gain from this.

3. There are other things that are still to be added to the Monte Carlo.
I wouldn't think that they will result in more than an additional 50 
ms/event.

Balancing all these factors I think the estimate above is reasonable.


Analysis:

1. The big uncertainty in the analysis has to do with the number of iterations
we need (once the TDC time is put in) for the Kalman filter to converge, as a
result of the drift distance dependence on the track angle. This should become
clearer in the next few months.

2. Once the timing is put in, there will be some addititonal calculations
in each iteration of the Kalman filter that will increase the CPU time.

3. The Kalman filter has not been optimized for CPU. I'm not sure at the moment
how much optimization can be done.

4. There are other calculations that will be done per event as we add more code
to the analysis.

These factors might suggest that we need more CPU time for the analysis, but
I think the number of iterations I assumed in this calculations is more than
we will actually need, so the estimated CPU time might turn out to be 
reasonable for the analysis as well.

Maher 
-- 
==================================================================
   Maher Quraan                        	   TRIUMF               
   TRIUMF TWIST Collaboration	           http://www.triumf.ca
   email: quraan@triumf.ca                 4004 Wesbrook Mall
   phone: (604)222-1047 ext. 6333	   Vancouver, BC
   fax:   (604)222-1074                    Canada V6T 2A3

		http://www.thehungersite.com/

==================================================================

Created for the The Center for Subatomic Research E614 Project Projects Page.

Created by The CoCoBoard.