d4.3

这是个trace drive的Cache模拟器

.RS
.IP \f3wback_always\fP 20n
Dirty data is always held in the cache, to be written back towards memory later.
.IP \f3wback_never\fP 20n
Dirty data is never held in the cache, i.e., this is a write-through policy.
.IP \f3wback_nofetch\fP 20n
Dirty data is held in the cache as long as no fetch is required.
A fetch would be required on a (sub-)block miss
if the write was not for an integral number of sub-blocks.
.IP \f3d4wback_impossible\fP 20n
This ``policy'' prints an error message and terminates the program;
it is for use only on read-only caches (e.g., instruction caches).
.RE
.IP \f3name_replacement\fP 15n
A pointer to a string describing the replacement policy.
This is for programmer use when printing results;
Dinero IV does nothing with it except require that it be
.RB non- NULL
when
.B d4setup
is called.
.IP \f3name_prefetch\fP 15n
A pointer to a string describing the prefetch policy.
This is for programmer use when printing results;
Dinero IV does nothing with it except require that it be
.RB non- NULL
when
.B d4setup
is called.
.IP \f3name_walloc\fP 15n
A pointer to a string describing the write allocate policy.
This is for programmer use when printing results;
Dinero IV does nothing with it except require that it be
.RB non- NULL
when
.B d4setup
is called.
.IP \f3name_wback\fP 15n
A pointer to a string describing the write back policy.
This is for programmer use when printing results;
Dinero IV does nothing with it except require that it be
.RB non- NULL
when
.B d4setup
is called.
.SH "D4CACHE RESULT FIELDS"
The result fields of each
.B d4cache
structure accumulate statistics, and are of primary interest
at the conclusion of simulation or after substantial amounts of simulation.
They are all of type
.BR double ,
because that has more precision than either integer or long
on most 32 bit machines.
They are all initialized to zero by
.BR d4new ,
and may be read or reset to zero by the user at any time.
.IP \f3multiblock\fP 18n
This field accumulates the total number of times
a reference affected more than one cache block.
Such references are split into two,
and this is done recursively as necessary,
so an original reference touching
.I n
cache blocks will ultimately cause
.B multiblock
to be incremented
.IR n \(mi1
times.
.IP \f3bytes_read\fP 18n
This field accumulates the total number of bytes read from downstream
(memory or the next larger cache).
.IP \f3bytes_written\fP 18n
This field accumulates the total number of bytes written downstream
(to memory or the next larger cache).
.PP
The following result fields are all arrays,
indexed by access type
(as described above, in
.BR "MEMORY REFERENCES" )
or an access type \(pl
.BR D4PREFETCH .
For example, to get the total number of misses,
one would add
.RS 18n
.sp \n(psu
.B miss[D4XREAD]
.br
\(pl
.B miss[D4XWRITE]
.br
\(pl
.B miss[D4XINSTRN]
.br
\(pl
.B miss[D4XMISC]
.br
\(pl
.B miss[D4XREAD\(plD4PREFETCH]
.br
\(pl
.B miss[D4XWRITE\(plD4PREFETCH]
.br
\(pl
.B miss[D4XINSTRN\(plD4PREFETCH]
.br
\(pl
.BR miss[D4XMISC\(plD4PREFETCH] .
.RE
.IP \f3fetch\fP 18n
These array values count the references processed for the cache.
.IP \f3miss\fP 18n
These array values count the cache misses.
.IP \f3blockmiss\fP 18n
These array values count the full cache block misses.
The difference between this array and the
.B miss
array is that
.B miss
also counts misses where only the sub-block actually referenced was
missing, while some other sub-blocks of the same block were valid in the cache.
.IP \f3comp_miss\fP 18n
These array values count the compulsory misses.
Compulsory misses are those that would occur even if the cache had infinite size.
The values in this array are not computed unless the
.B D4F_CCC
flag is set in
.BR flags .
.IP \f3comp_blockmiss\fP 18n
These array values count the compulsory full block misses.
Compulsory misses are those that would occur even if the cache had infinite size.
The values in this array are not computed unless the
.B D4F_CCC
flag is set in
.BR flags .
.IP \f3cap_miss\fP 18n
These array values count the capacity misses.
Capacity misses are those that would not occur if the cache had infinite size,
but would still occur if the cache was fully associative.
The values in this array are not computed unless the
.B D4F_CCC
flag is set in
.BR flags .
.IP \f3cap_blockmiss\fP 18n
These array values count the capacity full block misses.
Capacity misses are those that would not occur if the cache had infinite size,
but would still occur if the cache was fully associative.
The values in this array are not computed unless the
.B D4F_CCC
flag is set in
.BR flags .
.IP \f3conf_miss\fP 18n
These array values count the conflict misses.
Conflict misses are those that would not occur if the cache was fully associative.
The values in this array are not computed unless the
.B D4F_CCC
flag is set in
.BR flags .
.IP \f3conf_blockmiss\fP 18n
These array values count the conflict full block misses.
Conflict misses are those that would not occur if the cache was fully associative.
.B D4F_CCC
flag is set in
.BR flags .
.SH CUSTOMIZATION
Customization is a feature of Dinero IV that can offer significant
speedups for lengthy simulations.
It works by providing an easy way to recompile
the time critical cache simulation functions,
along the way replacing key
.B d4cache
structure field references by constants.
This allows partial evaluation and other optimizations
within the compiler to reduce time critical code path lengths.
The benefit varies according to cache configuration and simulated workload.
.PP
You use customization in the following way:
.IP 1. 4n
Your program performs initialization by calling
.B d4new
and
.BR d4setup ,
and directly modifying
.B d4cache
fields as explained in
.B DESCRIPTION
and
.BR "D4CACHE PARAMETER FIELDS" ,
above.
.IP 2. 4n
If the external constant integer
.B d4custom
is 0,
then call the function
.BI d4customize( F )\c
\&, where
.I F
is a
.IR stdio (3)
output stream open for writing.
The output file
will be written with C code containing customized functions
for the entire simulated memory hierarchy.
This code should be compiled and linked with the rest of the application
and the Dinero IV library to produce a new, customized, executable.
During compilation, the macro
.B D4CUSTOM
must be defined to 1 (typically by using the C compiler's
.B \-DD4CUSTOM
option).
.IP 3. 4n
The customized executable must start by performing exactly the
same initialization steps as the non-customized executable.
The external constant integer
.B d4custom
is 1.
.SH LIMITATIONS
The current version has no support for cache consistency,
and thus is of limited value for multiprocessor simulations.
.SH FILES
.B d4.h
\- header file.
.br
.B libd4.a
\- library.
.PP
While the Dinero IV header file and library may be installed anywhere,
we recommend keeping them with the Dinero IV source.
This means you will generally need to specify their location to the compiler.
The Dinero IV command implements customization by using the
.B D4_SRC
environment variable to locate the sources,
and we encourage use of the same convention in other programs.
.SH "SEE ALSO"
dineroIV (1).
.SH AUTHOR
Jan Edler and Mark D. Hill
([e-mail] and [e-mail], respectively).
.PP
The latest version of Dinero IV can be obtained from
.br
ftp://ftp.nj.nec.com/pub/edler/d4-\f2X\fP.tgz
.br
where \f2X\fP is the latest version number.
.SH COPYRIGHT
.PP
Copyright (C) 1997 NEC Research Institute, Inc. and Mark D. Hill.
.br
All rights reserved.
.br
Copyright (C) 1985, 1989 Mark D. Hill.  All rights reserved.
.PP
Permission to use, copy, modify, and distribute this software and
its associated documentation for non-commercial purposes is hereby
granted (for commercial purposes see below), provided that the above
copyright notice appears in all copies, derivative works or modified
versions of the software and any portions thereof, and that both the
copyright notice and this permission notice appear in the documentation.
NEC Research Institute Inc. and Mark D. Hill shall be given a copy of
any such derivative work or modified version of the software and NEC
Research Institute Inc.  and any of its affiliated companies (collectively
referred to as NECI) and Mark D. Hill shall be granted permission to use,
copy, modify, and distribute the software for internal use and research.
The name of NEC Research Institute Inc. and its affiliated companies
shall not be used in advertising or publicity related to the distribution
of the software, without the prior written consent of NECI.  All copies,
derivative works, or modified versions of the software shall be exported
or reexported in accordance with applicable laws and regulations relating
to export control.  This software is experimental.  NECI and Mark D. Hill
make no representations regarding the suitability of this software for
any purpose and neither NECI nor Mark D. Hill will support the software.
.PP
Use of this software for commercial purposes is also possible, but only
if, in addition to the above requirements for non-commercial use, written
permission for such use is obtained by the commercial user from NECI or
Mark D. Hill prior to the fabrication and distribution of the software.
.PP
THE SOFTWARE IS PROVIDED AS IS.  NECI AND MARK D. HILL DO NOT MAKE
ANY WARRANTEES EITHER EXPRESS OR IMPLIED WITH REGARD TO THE SOFTWARE.
NECI AND MARK D. HILL ALSO DISCLAIM ANY WARRANTY THAT THE SOFTWARE IS
FREE OF INFRINGEMENT OF ANY INTELLECTUAL PROPERTY RIGHTS OF OTHERS.
NO OTHER LICENSE EXPRESS OR IMPLIED IS HEREBY GRANTED.  NECI AND MARK
D. HILL SHALL NOT BE LIABLE FOR ANY DAMAGES, INCLUDING GENERAL, SPECIAL,
INCIDENTAL, OR CONSEQUENTIAL DAMAGES, ARISING OUT OF THE USE OR INABILITY
TO USE THE SOFTWARE.