Thank you for your reply...
Given your info, applu takes 10807886 dynamic instructions and simple
atomic CPU should have 1.6 ~ 2.3 M instruction / s simulation speed, it
only will take 1 minute to simulate applu. However, on my machine and M5
I have, it has taken *1 hour* and still running...
The compiler I used is from gcc 4.3.2 available on M5 wiki:
alphaev67-unknown-linux-gnu/bin/alphaev67-unknown-linux-gnu-gfortran
The cmd line I used: ./m5.fast ../../configs/Test_Weixun/se.py -c
../../../benchmarks/cpu2000/applu/applu -i
../../../benchmarks/cpu2000/applu/applu.in
Here I used the input file given in data/ref/input/applu.in which is
less than 1K. Which reduced-size input files you used?
AtomicSimpleCPU and atomic mem access are used
I checked ps -u and found m5.fast is taking 100% CPU and is doing
something....
[wewang 3877 99.9 1.3 580396 216800 pts/3 R<+ 13:29 50:16
./m5.fast ../.]
Please help.
-Weixun
On 6/3/2010 4:44 PM, soumyaroop roy wrote:
Which cpu are you using? Simple atomic?
I verified that all the integer spec2000 benchmarks work correctly
(most of them were run using smred input sets) with simple atomic, o3,
and in-order. Here are some numbers from some runs that I did in Feb,
2010 (format - benchmark name : dynamic instruction count) for
simple-atomic.
applu 10807886
sixtrack 11730403
mgrid 15410653
vpr (vpr_place) 17615460
lucas 19239685
swim 23563883
galgel 33656971
gap 44751803
ammp 45811363
facerec 63152814
apsi 71271052
vortex 88340673
twolf 91902974
eon (eon_cook) 94039676
crafty 94420287
gcc 96772979
mcf 188605120
perlbmk (perlbmk_makerand) 200610007
parser 267811614
gzip (gzip_log) 601856964
equake 1021602432
mesa 1608475031
art 1660422322
bzip2 (bzip2_source) 1819780127
wupwise seg fault
fma3d seg fault
The machines (standard x86_64 desktops/servers) that I ran my
simulations on, simple-atomic would yield a simulation speeds of 1.6 -
2.3 M instructions/second. O3 was about 0.03X and inorder was 0.01X.
So, that should give you a rough estimate of the runtimes that you
should be expecting at you end. But if you do not use the reduced
input sets, then it may take forever.
I never ran any of my simulations by specifying options from the
command line. I created mibench.py and mediabench.py like cpu2000.py
to set up simulations for these benchmarks. There's a bug in the
cpu2000.py currently (at least, there was one till Nov/Dec 2009) that
break the output redirection. So, I just had to fix that.
Also, if the licensing of mediabench and mibench allow distributing
binaries, I'd be happy to contribute them (alpha binaries currently)
along with the benchmark configuration scripts.
regards,
Soumyaroop
On Thu, Jun 3, 2010 at 4:15 PM, Weixun Wang<[email protected]> wrote:
I'm usinig m5.fast now. Everything is OK except the following:
1) Benchmarks from CPU 2000 are still executing very slow in M5 (m5.fast). I
tried gcc and mcf. Both of them have taken quite a long time but are still
running. I'm sure they are being executed based on the output. Is it normal?
2) MiBench benchmarks work well. However, I found output redirection does
not work. For example, running toast:
./m5.fast ../../configs/Test_Weixun/se.py --cmd
../../../benchmarks/mibench/gsm/toast --options "-fps -c
../../../benchmarks/mibench/gsm/small.au>
../../../benchmarks/mibench/gsm/output_small.encode.gsm"
The redirection part of --options ">
../../../benchmarks/mibench/gsm/output_small.encode.gsm" does not have any
effect and the output is to the screen no matter what. How to specify
redirection in this scenario?
3) gzip from CPU 2000 gives the following error after executing for a
certain amount of time:
panic: Tried to access unmapped address 0x12014a000.
@ cycle 141067250000
[invoke:build/ALPHA_SE/arch/alpha/faults.cc, line 204]
Memory Usage: 576044 KBytes
For more information see: http://www.m5sim.org/panic/5932f339
Program aborted at cycle 141067250000
Abort
I havn't found other benchmarks which give errors during simulation.
-Weixun
On 6/3/2010 12:04 PM, soumyaroop roy wrote:
Are you dumping any traces? What stops you from using m5.fast?
-Soumyaroop
On Thu, Jun 3, 2010 at 11:37 AM, Weixun Wang<[email protected]> wrote:
I used the pre-compiled gcc 3.4.3 available on M5 webpage.
BTW, will m5.opt be much more faster?
Thanks.
-Weixun
On 6/3/2010 11:17 AM, soumyaroop roy wrote:
You may try dijkstra with the small input set. It is very short.
Which version of GCC are you using?
-Soumyaroop
On Thu, Jun 3, 2010 at 11:01 AM, Weixun Wang<[email protected]>
wrote:
Hi List,
Now I compiled MiBench as well as CPU 2000 (both have to be one
benchmark
by
one benchmark) for M5, using options simply "-static -o XXX -lm"
removing
"-O3", and found the compiled binaries executing in M5 properly. Lucky
me...
However, I also found those benchmarks running, especially CPU 2000
like
gzip, extremely slow for m5.debug. I truncated the input file for gzip
to
only 34K but the simulation still takes forever...
-Weixun
On 6/3/2010 10:16 AM, soumyaroop roy wrote:
The makefiles I used for MiBench and MediaBench are nothing special.
They should work just fine. As Max mentioned earlier, you just have to
make sure that you create statically linked binaries.
One thing I noticed is that you're using an older GCC (3.4.3). Did you
try GCC 4.3.2?
-Soumyaroop
On Thu, Jun 3, 2010 at 10:00 AM, Weixun Wang<[email protected]>
wrote:
Hey Max,
As you said, I did put -static. However, I'm not sure whether "-im"
is
appropriate in this scenario.
If anyone has successfully compiled MiBench or CPU 2000, could you
share
your Makefile with me?
Thanks.
Weixun
On 6/3/2010 3:16 AM, Maximilien Breughe wrote:
Hey,
1) Did you compile the benchmarks statically? Because if you want to
run
them on M5 everything you run is the compiled binary. So dynamic
libraries
cannot be load.
2) For SPEC CPU2000 the situation is indeed more complicated. I
compiled
some benchmarks of SPEC CPU2006 and had to do the following things in
the
config-file (make sure you pick a config-file with properties very
similar
to your machine):
-Select the propper compiler:
CC = gcc
into
CC = /DIRECTORY OF
CROSSTOOL/gcc-3.4.3-glibc-2.3.5/alpha-unknown-linux-gnu/bin/alpha-unknown-linux-gnu-gcc
-Modify the compile-flags (add the "static" flag) in the
"optimization"
section:
## Base is low opt
default=base=default=default:
COPTIMIZE = -O2 -static
CXXOPTIMIZE = -O2 -static
FOPTIMIZE = -O2 -static
However I recommend you to build SPEC CPU 2000 first for your native
machine. This could be already difficult. If it works, then try it
for
ALPHA_SE
Greets,
Max
On 06/02/2010 10:25 PM, Weixun Wang wrote:
Hi all,
I want to cross-compile some benchmarks so that I can run them in M5.
I'm
using a ALPHA_SE mode M5 on a x86_64 Ubuntu machine. Potentially I
target
to
use MiBench or SPEC CPU 2000. I downloaded the pre-compiled
cross-compiler
for Alpha from m5sim.org.
1) For MiBench, it has a simple src code/data structure with a simple
Makefile. For example, qsort, I modifed the Makefile of it as
follows:
-------------------------------------------------------------------------------------------------
FILE1 = qsort_small.c
FILE2 = qsort_large.c
all: qsort_small qsort_large
qsort_small: qsort_small.c Makefile
/export/research57/weixun/tools/M5/crosstool/gcc-3.4.3-glibc-2.3.5/alpha-unknown-linux-gnu/bin/alpha-unknown-linux-gnu-gcc
-static qsort_small.c -O3 -o qsort_small -lm
qsort_large: qsort_large.c Makefile
/export/research57/weixun/tools/M5/crosstool/gcc-3.4.3-glibc-2.3.5/alpha-unknown-linux-gnu/bin/alpha-unknown-linux-gnu-gcc
-static qsort_large.c -O3 -o qsort_large -lm
clean:
rm -rf qsort_small qsort_large output*
-------------------------------------------------------------------------------------------------
I used the option "-lm" since the original "-lc" lead to errors like
"undefined references to 'exp'". If "-L
/export/research57/weixun/tools/M5/crosstool/gcc-3.4.3-glibc-2.3.5/alpha-unknown-linux-gnu/lib"
is used, the error still remains.
The compilation is successful. However, if I run it using
configs/example/se.py, I got the errors:
-------------------------------------------------------------------------------------------------
.........................................................
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
info: Increasing stack size by one page.
panic: Tried to access unmapped address 0xa8.
@ cycle 4947000
[invoke:build/ALPHA_SE/arch/alpha/faults.cc, line 204]
Memory Usage: 577796 KBytes
For more information see: http://www.m5sim.org/panic/5932f339
Program aborted at cycle 4947000
Abort
-------------------------------------------------------------------------------------------------
Other benchmarks in MiBench shows similar errors or just simplely
hang
on
"info: Increasing stack size by one page.".
I guess the compile options are not correct or the right library is
not
linked. Please help...........................
2) For SPEC CPU 2000, it is more complicated. I copied the entire CD
to
a
local directory. The problem is I do not have the su privilege on the
machine I worked on. So the install.sh does not work correctly. Is
there
a
simpler way for cross-compile SPEC CPU 2000 benchmarks (or at least
some
of
them) for M5?
Thanks!!!
--
Best Regards,
Wang, Weixun
Department of Computer& Information Science& Engineering
Gator College of Engineering
University of Florida
Gainesville, FL 32611
http://www.cise.ufl.edu/~wewang
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
--
Best Regards,
Wang, Weixun
Department of Computer& Information Science& Engineering
Gator College of Engineering
University of Florida
Gainesville, FL 32611
http://www.cise.ufl.edu/~wewang
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
--
Best Regards,
Wang, Weixun
Department of Computer& Information Science& Engineering
Gator College of Engineering
University of Florida
Gainesville, FL 32611
http://www.cise.ufl.edu/~wewang
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
--
Best Regards,
Wang, Weixun
Department of Computer& Information Science& Engineering
Gator College of Engineering
University of Florida
Gainesville, FL 32611
http://www.cise.ufl.edu/~wewang
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
--
Best Regards,
Wang, Weixun
Department of Computer& Information Science& Engineering
Gator College of Engineering
University of Florida
Gainesville, FL 32611
http://www.cise.ufl.edu/~wewang
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
--
Best Regards,
Wang, Weixun
Department of Computer& Information Science& Engineering
Gator College of Engineering
University of Florida
Gainesville, FL 32611
http://www.cise.ufl.edu/~wewang
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
