Index: trunk/doc/design/ippSDRS.tex
===================================================================
--- trunk/doc/design/ippSDRS.tex	(revision 2176)
+++ trunk/doc/design/ippSDRS.tex	(revision 2186)
@@ -1,3 +1,3 @@
-%%% $Id: ippSDRS.tex,v 1.9 2004-10-20 08:17:36 eugene Exp $
+%%% $Id: ippSDRS.tex,v 1.10 2004-10-21 03:55:59 eugene Exp $
 \documentclass[panstarrs]{panstarrs}
 
@@ -2370,28 +2370,96 @@
 \section{Computer Hardware}
 
-\subsection{PS-1 Cluster requirements}
-
-  \begin{itemize}
-    \item CPU requirements
-    \item per-node I/O requirements
-    \item switch throughput requirements
-    \item storage profile
-  \end{itemize}
-
-\subsection{PS-1 Cluster Hardware Plan}
-
-  \begin{itemize}
-    \item COTS equipment
-    \item number of processors needed
-    \item number of I/O ports needed
-    \item number of disk slots needed
-    \item switch choice
-    \item design choice for computer nodes
-    \item total rack space
-  \end{itemize}
+\subsection{PS-1 Cluster Design}
+  
+The PS-1 IPP computer system is designed as a cluster of 'fat bricks':
+computers with both processing power and large amounts of local disk
+storage.  These computers are large rack-mount boxes with space for
+10s of disks (24 and 36 disk cases are available) and a motherboard
+with two CPUs and two Gig-E ethernet ports.  One set of machines is
+specified for storage and processing of the individual OTAs up through
+Phase 2 (the `OTA nodes'), another set of machines are specified for
+storage of the Static Sky and processing of data from Phase 3 and
+Phase 4 (the `Sky nodes').  Other machines will be necessary to
+support the Metadata DB and the AP DB.  
+
+The IPP PS-1 SRS (PSDC-xxx) specifies the processing throughput
+requirements for the IPP.  We have performed benchmark tests of the
+processing needs in order to achieve this throughput.  The details of
+this study are presented in the IPP Hardware Analysis (PSDC-xxx),
+which we summarize here.  The analysis measures the processing time
+(excluding I/O) for both Phase 2 and Phase 4 on an Intel Pentium 4
+processor, and expresses the processing time in GHz-seconds, under the
+assumption that a machine with the same architecture and twice the
+processor speed with perform the same analysis in half the time.  This
+is probably a valid assumption within a limited range on hardware
+using the same architecture.  We independently find that 32-bit Pentium
+processors perform somewhat slower (up to a factor of 2) than
+equivalently rated 64 bit Opeteron processors.  This discrepancy makes
+our numbers somewhat conservative, but may only compensate for the
+simplistic analysis we have performed.  
+
+Our benchmarks show that the Phase 2 analysis takes 12000 GHz-seconds
+for a complete major frame (4 FPAs) while the Phase 4 analysis takes
+7800 GHz-seconds for the same major frame.  We also examine the total
+data I/O required for each processing node both locally to disk and
+across the network to other machines.  These numbers in turn depend on
+whether the data is optimally stored on the OTA nodes (raw images
+matched to their calibration images) or if the data are randomized.
+There are also differences in the analysis for how many bits and
+images are used in the processing.  For PS-1, the `minimal' data set
+is approrpiate, resulting in a total Phase 2 I/O of 21 GBs per major
+frame and a total Phase 4 I/O of 36 GBs.  We will use the randomized
+numbers as a conservative estimate, and assume the network is the
+dominant I/O bottleneck.
+
+The analysis assumes each CPU is associated with one RAID array
+(maximum throughput 110 MB/sec) and one network controller (maximum
+throughput 70 MB/s) and that each one is a 2.2 GHz processor. In this
+case, given the CPU load and I/O throughput above, the Phase 2 will
+require a total of 190 seconds of I/O and 5500 seconds of processing
+distributed across the cluster.  Likewise, the Phase 4 analysis will
+require a total of 330 sec of I/O and 3500 seconds of processing.
+Given the 160 seconds available per major frame, these numbers imply a
+total of 63 processors are needed to keep up with the processing and
+I/O load.  
+
+The other major driver on the IPP PS-1 cluster are the data storage
+requirements.  We are required to store the entire AP Survey data and
+the IVP data, and to have storage enough to represent the Static Sky
+by the end of the two year mission.  These storage requirements as a
+function of time are shown in Figure~\ref{StorageProfile}.  Based on
+the PS-1 Design Reference Mission (PSDC-xxx), by the end of the
+second year, we will have total storage needs of 850 TB for raw images
+and the Static Sky, and an additional \tbd{XXX} TB for the AP DB
+storage.  
+
+To meet these requirements, we have designed the IPP cluster to use
+fat bricks which will be capable of holding 24 disks each.  Before
+PS-1 goes on line, we will purchase enough disks to fill 1/3 of the
+disk slots.  After 9 months (2006 Sept), we will purchase the next 1/3
+of the disks, and the remaining disks 9 months after that (2007 June).
+We have made conservative estimates of the available disk sizes at
+these purchase dates (400 GB, 600 GB, and 900 GB), allowing us to
+determine the number of computers needed to meet the storage
+specification.  We will purchase 80 computers, with the storage
+profile shown in the figure, ending at a total capacity (after
+discounting volume for RAID overhead and binary vs digital terabytes)
+of 950 TBs.  The 80 computers will easily meet the processing and I/O
+requirements given the above need to 63 processors.  
+
+\tbd{switch information}
+
+\tbd{RAID and compression / duplication plan}
 
 \subsection{PS-1 Cluster Expected Reliability}
   
 \subsection{PS-1 Cluster Support}
+
+\begin{figure}
+\begin{center}
+\resizebox{6in}{!}{\includegraphics{pics/ps1_ipp_storage.ps}}
+\caption{ \label{StorageProfile} Storage Profile}
+\end{center}
+\end{figure}
 
 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%