%* Last edited: Feb 13 14:18 1995 (misha)

\section{Notes (ISCA)}

\paragraph{Working Sets, Cache Sizes, Node Granularity Issues}

Rothberg, Sing, Gupta.

\begin{enumerate}
\item Analyzed applications:  fundamental working set characteristics.
\item Found sharply defined ``knees''
\item Used high radix FFTs to reduce communication.
\item Some applications not graint-size scalable (?)
\item Applications: block LU decomp all use 4-8K cache.
\end{enumerate}

\paragraph{Chinese Remainder Theorem}

Look up  reference A. Aviegensis, Transaction Computers, C-20, Nov '71.
Illiac-IV, Sep 76, Trans Comp, C-25.

I forget why this is interesting.

\paragraph{Kuck Talk}

\begin{itemize}

\item Divided history into architecture to compiler to applications to problem
solving.

\item Speed estimates of parallel machines were always optimistic.

\item General rule, 10x performance in 7 years. Therefore 100 in 15 years!
  Real example: Cray 1 to C90/8.  15 years: Clock 3x, Peak 50x.


\item 8086/Pentium example.  Clock speed 66 vs 4.77 (138x) but performance
  increased from .33 to 112 (336x). Rest came from architecture and
  compilers.

\end{itemize}