An ARM comparison: Raspberry Pi vs OpenRD Client

Bellow there’s a comparison between Raspberry Pi (Model B) and OpenRD Client, two low-power ARM SoC (System on a Chip), running Linux (Rasbian and Fedora respectively). I used the EEMBC CoreMark benchmark with default settings for 100000, 1000000 and 2000000 iterations. For each set, each device keep its performance with the same rate of iterations per second. The difference between the devices is at, an average level of, 520 iterations per second. That translates to 1.4x faster in favour of the OpenRD Client.

While OpenRD Client performs better, in terms of efficiency, Raspberry Pi seems to doing much better. In terms of CPU, Raspbeery Pi uses a 700MHz (with the possibility of setting it up to 1GHz) ARMv11 processor while OpenRD Client an older 1.2GHz Marvell Sheeva (ARMv5TE). ARMv11 supports an FPU unlike ARMv5 and that is one of the reasons that Raspberry Pi performs quite well despite its lower specifications. In terms of memory, Raspberry Pi uses 256MB DDR2 800MHz (newer models have 512MB RAM), in share with its relatively powerful GPU, while OpenRD Client uses 512MB DDR2 800MHz. OpenRD Client also provides Gigabit Ethernet as well as many more I/O interfaces. Also, OpenRD has preloaded OS, unlike Raspberry Pi that boots from an SD card. During this benchmark, Raspberry Pi was running Rasbian “Wheezy” (unofficial Debian port to armhf) while OpenRD Client Fedora 8.

(Raspberry Pi diagram – http://www.raspberrypi.org/wp-content/uploads/2012/04/Raspi_Iso_Blue.png)

(OpenRD Client Hardware Block diagram – http://www.globalscaletechnologies.com/skins/skin_1/images/OpenRD-Client_back.PNG)

Advertisement

Size matters

…for power consumption.

OpenRD Client with Marvel Sheeva 88F6281 (7 Watt idle, single core)

Intel Atom D525 on Mini-ITX bundle (45 Watt idle, dual-core hyper-threading).

Dell PowerEdge r610 with Intel Xeon E5507 (119 Watt idle, double quad-core)

Compiling SPEC Benchmarks tools on ARM

Some of the SPEC benchmarks come for a variety of architectures but not for ARM, which is not surprising anyway. The easiest way to execute the benchmarks is by using the provided scripts, which is pretty much straight forward operation. The benchmarks come with pre-compiled tools and libraries that are needed by the benchmarks. Among them is Perl, which will fail to compile on ARM systems because of some invalid object definitions in its makefile. These definitions get in the makefile by the Configure script which is called before calling make. Instead of messing around with Configure, I added a couple of lines in the buildtools file, which is responsible for building the SPEC tools on a new architecture or after having modified the host system, in order to remove the unnecessary lines from the corresponding makefiles.

The buildtools script can be found under tools/src/ in the SPEC benchmarks directoy. What is needed is to go in the Perl building section, line 103, and replace the fist part of the building phase with the following oneliner:

(cd $PERLSRC; LD_LIBRARY_PATH=`pwd`; export LD_LIBRARY_PATH; 
./Configure -dOes $PERLFLAGS -Ddosuid=undef -Dprefix=$INSTALLDIR 
-Dd_bincompat3=undef; cat makefile | grep -v built-in 
| grep -v "command line" > makefile.new; cp makefile.new makefile; 
cat x2p/makefile | grep -v built-in | grep -v "command line" > x2p/makefile.new; 
cp x2p/makefile.new x2p/makefile; make; ./perl installperl ) 
|| die "Can't build perl"

Running Fortran on ARM

Generally and very briefly, there is no official Fortran compiler available for the ARM architecture. The easiest way to get Fortran code (specifically Fortran 77) to run on the ARM architecture, or any other architecture that doesn’t have a Fortran compiler, is to convert the Fortran code to C. In order to do that efficiently, we can use Netlib’s f2c command line tool, available for Linux, Unix and Windows.

The steps that need to be followed are:

1) Compile the library – source at http://www.netlib.org/f2c/libf2c.zip
Rename makefile.u to makefile and do make. Copy the generated libf2c.a to /usr/lib and the f2c.h header under /usr/include.
2) Compile the binary – source at http://www.netlib.org/f2c/src/
Rename makefile.u to makefile and do make. Copy the binary f2c under /usr/local/bin.

Supposing we have the foo.f file we can generate the C version by running f2c foo.f, giving us foo.c. It can now be compiled using ‘gcc foo.c -o foo -lf2c‘.

In order to automate this process a while, the following script can be used. It accepts only one argument, the Fortran source file.

#!/bin/bash
fortranFile=$1
fileName=`echo $1 | sed 's/\(.*\)\..*/\1/'`
echo $fileName
f2c $fortranFile
gcc ${fileName}.c -o $fileName -lf2c

UPDATE: There is GCC Fortran compiler for ARM Fedora and Debian and I presume for other distros as well. The issue now is that these distributions are compiled for ARMv5, while the latest ARM processors (Cortex-A8, A9, A15) are of the ARMv7 architecture. The compilers, and the OS in general is therefore unable to make use of the additional instructions sets and FPU. Other distros, such as Slackware, are compiled on even older architecture, ARMv4.