SBCX-TN-006: Characterizing the RAM bandwidth of Axel Lite SoM

Info Box Applies to SBC AXEL Applies to Axel Ultra Applies to Axel Lite Applies to AXEL ESATTA

This technical note was validated against specific versions of hardware and software. What is described here may not work with other versions.

History

Version	Date	Notes
1.0.0	January 2020	First public release

Introduction

When dealing with computationally heavy tasks, RAM bandwidth factor may turn out to be a severe bottleneck bounding the overall performance, especially in regard to the SoC's used to build embedded systems. For this reason, characterizing RAM bandwidth is useful when dealing with such applications.

This technical note (TN for short) illustrates several benchmarking tests that were run on Axel Lite SoM to characterize this bandwidth. This SoM is built upon the i.MX6Q/D/DL/S family of processors by NXP.

Testbed general configuration

This section illustrates the configuration settings common to all the tests that were performed. Basically, the testbed is the same described in this TN.

SoC and SDRAM bank

The SoC model is i.MX8M Quad:

armbian@sbcx:~/devel/stream/lmbench$ lscpu 
Architecture:        armv7l
Byte Order:          Little Endian
CPU(s):              4
On-line CPU(s) list: 0-3
Thread(s) per core:  1
Core(s) per socket:  4
Socket(s):           1
Vendor ID:           ARM
Model:               10
Model name:          Cortex-A9
Stepping:            r2p10
CPU max MHz:         996.0000
CPU min MHz:         396.0000
BogoMIPS:            7.54
Flags:               half thumb fastmult vfp edsp neon vfpv3 tls vfpd32

This processor is capable of running at different speeds. All the tests were conducted at 996 MHz.

The following table details the characteristics of the SDRAM bank connected to the SoC.

		Platform
		Mito8M	AxelLite
SoC	SoC	NXP i.MX8M Quad
	ARM frequency [MHz]	800 or 1300
SDRAM	Type	LPDDR4
	Frequency [MHz]	1600
	Bus witdth [bit]	32
	Theoretical bandwidth [Gb/s]	102.4
	Theoretical bandwidth [GB/s]	12.8
	Size [MB]	3072

Software configuration

• Linux kernel: 4.14.98
• Root file system: Debian GNU/Linux 10 (buster)
• Architecture: aarch64
• Governor: userspace @ 800 MHz

root@Mito8M:~# echo userspace > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
root@Mito8M:~# cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor
userspace
root@Mito8M:~# cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_cur_freq
800000

As some benchmarks were built natively on the platform under test itself, the version of the GCC compiler is indicated as well:

armbian@Mito8M:~/devel/lmbench$ gcc -v
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/lib/gcc/aarch64-linux-gnu/8/lto-wrapper
Target: aarch64-linux-gnu
Configured with: ../src/configure -v --with-pkgversion='Debian 8.3.0-6' --with-bugurl=file:///usr/share/doc/gcc-8/README.Bugs --enable-languages=c,ada,c++,go,d,fortran,objc,obj-c++ --prefix=/usr --with-gcc-major-version-only --program-suffix=-8 --program-prefix=aarch64-linux-gnu- --enable-shared --enable-linker-build-id --libexecdir=/usr/lib --without-included-gettext --enable-threads=posix --libdir=/usr/lib --enable-nls --enable-bootstrap --enable-clocale=gnu --enable-libstdcxx-debug --enable-libstdcxx-time=yes --with-default-libstdcxx-abi=new --enable-gnu-unique-object --disable-libquadmath --disable-libquadmath-support --enable-plugin --enable-default-pie --with-system-zlib --disable-libphobos --enable-multiarch --enable-fix-cortex-a53-843419 --disable-werror --enable-checking=release --build=aarch64-linux-gnu --host=aarch64-linux-gnu --target=aarch64-linux-gnu
Thread model: posix
gcc version 8.3.0 (Debian 8.3.0-6)

Benchmarks

STREAM

TBD

LMbench

TBD

pmbw

TBD

Overall results

This section details the results that were achieved by the different benchmarks

STREAM

		Mito8M
	ARM frequency [MHz]	792
	Frequency [MHz]	1600
	Bus witdth [bit]	32

LMbench

TBD

pmbw

TBD

Useful links

• STREAM benchmark
• LM Bench benchmark
• pmbw benchmark
• Joshua Wyatt Smith and Andrew Hamilton, Parallel benchmarks for ARM processors in the highenergy context
• T Wrigley, G Harmsen and B Mellado, Memory performance of ARM processors and itsrelevance to High Energy Physics
• G. T. Wrigley, R. G. Reed, B. Mellado, Memory benchmarking characterisation of ARM-based SoCs

Appendix A: Detailed testing procedures

This section details how the benchmarks were configured and run on the testbed.

STREAM

Building

git clone https://github.com/jeffhammond/STREAM.git
make

 1 armbian@Mito8M:~/devel/STREAM$ cat Makefile 
 2 CC = gcc
 3 CFLAGS = -O2 -fopenmp
 4 
 5 FC = gfortran-4.9
 6 FFLAGS = -O2 -fopenmp
 7 
 8 all: stream_c.exe
 9 
10 stream_f.exe: stream.f mysecond.o
11         $(CC) $(CFLAGS) -c mysecond.c
12         $(FC) $(FFLAGS) -c stream.f
13         $(FC) $(FFLAGS) stream.o mysecond.o -o stream_f.exe
14 
15 stream_c.exe: stream.c
16         $(CC) $(CFLAGS) stream.c -o stream_c.exe
17 
18 clean:
19         rm -f stream_f.exe stream_c.exe *.o
20 
21 # an example of a more complex build line for the Intel icc compiler
22 stream.icc: stream.c
23         icc -O3 -xCORE-AVX2 -ffreestanding -qopenmp -DSTREAM_ARRAY_SIZE=80000000 -DNTIMES=20 stream.c -o stream.omp.AVX2.80M.20x.icc

Running the tests

armbian@sbcx:~/devel/stream/STREAM$ ./stream_c.exe 
-------------------------------------------------------------
STREAM version $Revision: 5.10 $
-------------------------------------------------------------
This system uses 8 bytes per array element.
-------------------------------------------------------------
Array size = 10000000 (elements), Offset = 0 (elements)
Memory per array = 76.3 MiB (= 0.1 GiB).
Total memory required = 228.9 MiB (= 0.2 GiB).
Each kernel will be executed 10 times.
 The *best* time for each kernel (excluding the first iteration)
 will be used to compute the reported bandwidth.
-------------------------------------------------------------
Number of Threads requested = 4
Number of Threads counted = 4
-------------------------------------------------------------
Your clock granularity/precision appears to be 1 microseconds.
Each test below will take on the order of 153104 microseconds.
   (= 153104 clock ticks)
Increase the size of the arrays if this shows that
you are not getting at least 20 clock ticks per test.
-------------------------------------------------------------
WARNING -- The above is only a rough guideline.
For best results, please be sure you know the
precision of your system timer.
-------------------------------------------------------------
Function    Best Rate MB/s  Avg time     Min time     Max time
Copy:            1139.8     0.141910     0.140376     0.144318
Scale:           1124.8     0.143720     0.142245     0.144615
Add:             1185.1     0.204792     0.202517     0.206718
Triad:           1214.4     0.201321     0.197631     0.202673
-------------------------------------------------------------
Solution Validates: avg error less than 1.000000e-13 on all three arrays
-------------------------------------------------------------

LMbench

Running the tests

armbian@sbcx:~/devel/stream/STREAM$ sudo lmbench-run 
=====================================================================

                L M B E N C H   C ON F I G U R A T I O N
                ----------------------------------------

You need to configure some parameters to lmbench.  Once you have configured
these parameters, you may do multiple runs by saying

        "make rerun"

in the src subdirectory.

NOTICE: please do not have any other activity on the system if you can
help it.  Things like the second hand on your xclock or X perfmeters
are not so good when benchmarking.  In fact, X is not so good when
benchmarking.

=====================================================================

If you are running on an MP machine and you want to try running
multiple copies of lmbench in parallel, you can specify how many here.

Using this option will make the benchmark run 100x slower (sorry).

NOTE:  WARNING! This feature is experimental and many results are 
        known to be incorrect or random!

MULTIPLE COPIES [default 1]: 
=====================================================================

Options to control job placement
1) Allow scheduler to place jobs
2) Assign each benchmark process with any attendent child processes
   to its own processor
3) Assign each benchmark process with any attendent child processes
   to its own processor, except that it will be as far as possible
   from other processes
4) Assign each benchmark and attendent processes to their own
   processors
5) Assign each benchmark and attendent processes to their own
   processors, except that they will be as far as possible from
   each other and other processes
6) Custom placement: you assign each benchmark process with attendent
   child processes to processors
7) Custom placement: you assign each benchmark and attendent
   processes to processors

Note: some benchmarks, such as bw_pipe, create attendent child
processes for each benchmark process.  For example, bw_pipe
needs a second process to send data down the pipe to be read
by the benchmark process.  If you have three copies of the
benchmark process running, then you actually have six processes;
three attendent child processes sending data down the pipes and 
three benchmark processes reading data and doing the measurements.

Job placement selection [default 1]: 
=====================================================================

Hang on, we are calculating your timing granularity.
OK, it looks like you can time stuff down to 5000 usec resolution.

Hang on, we are calculating your timing overhead.
OK, it looks like your gettimeofday() costs 0 usecs.

Hang on, we are calculating your loop overhead.
OK, it looks like your benchmark loop costs 0.00000055 usecs.

=====================================================================

Several benchmarks operate on a range of memory.  This memory should be
sized such that it is at least 4 times as big as the external cache[s]
on your system.   It should be no more than 80% of your physical memory.

The bigger the range, the more accurate the results, but larger sizes
take somewhat longer to run the benchmark.

MB [default 1186]: 1024
Checking to see if you have 1024 MB; please wait for a moment...
1024MB OK
1024MB OK
1024MB OK
Hang on, we are calculating your cache line size.
OK, it looks like your cache line is 32 bytes.

=====================================================================

lmbench measures a wide variety of system performance, and the full suite
of benchmarks can take a long time on some platforms.  Consequently, we
offer the capability to run only predefined subsets of benchmarks, one
for operating system specific benchmarks and one for hardware specific
benchmarks.  We also offer the option of running only selected benchmarks
which is useful during operating system development.

Please remember that if you intend to publish the results you either need
to do a full run or one of the predefined OS or hardware subsets.

SUBSET (ALL|HARWARE|OS|DEVELOPMENT) [default all]: 
=====================================================================

This benchmark measures, by default, memory latency for a number of
different strides.  That can take a long time and is most useful if you
are trying to figure out your cache line size or if your cache line size
is greater than 128 bytes.

If you are planning on sending in these results, please don't do a fast
run.

Answering yes means that we measure memory latency with a 128 byte stride.  

FASTMEM [default no]: 
=====================================================================

This benchmark measures, by default, file system latency.  That can
take a long time on systems with old style file systems (i.e., UFS,
FFS, etc.).  Linux' ext2fs and Sun's tmpfs are fast enough that this
test is not painful.

If you are planning on sending in these results, please don't do a fast
run.

If you want to skip the file system latency tests, answer "yes" below.

SLOWFS [default no]: yes
=====================================================================

This benchmark can measure disk zone bandwidths and seek times.  These can
be turned into whizzy graphs that pretty much tell you everything you might
need to know about the performance of your disk.  

This takes a while and requires read access to a disk drive.  
Write is not measured, see disk.c to see how if you want to do so.

If you want to skip the disk tests, hit return below.

If you want to include disk tests, then specify the path to the disk
device, such as /dev/sda.  For each disk that is readable, you'll be
prompted for a one line description of the drive, i.e., 

        Iomega IDE ZIP
or
        HP C3725S 2GB on 10MB/sec NCR SCSI bus

DISKS [default none]: 
=====================================================================

If you are running on an idle network and there are other, identically
configured systems, on the same wire (no gateway between you and them),
and you have rsh access to them, then you should run the network part
of the benchmarks to them.  Please specify any such systems as a space
separated list such as: ether-host fddi-host hippi-host.

REMOTE [default none]: 
=====================================================================

Calculating mhz, please wait for a moment...
I think your CPU mhz is 

        1992 MHz, 0.5020 nanosec clock

but I am frequently wrong.  If that is the wrong Mhz, type in your
best guess as to your processor speed.  It doesn't have to be exact,
but if you know it is around 800, say 800.  

Please note that some processors, such as the P4, have a core which
is double-clocked, so on those processors the reported clock speed
will be roughly double the advertised clock rate.  For example, a
1.8GHz P4 may be reported as a 3592MHz processor.

Processor mhz [default 1992 MHz, 0.5020 nanosec clock]: 996
=====================================================================

We need a place to store a 1024 Mbyte file as well as create and delete a
large number of small files.  We default to /var/tmp.  If /var/tmp is a
memory resident file system (i.e., tmpfs), pick a different place.
Please specify a directory that has enough space and is a local file
system.

FSDIR [default /var/tmp/lmbench]: /home/armbian/devel/stream/lmbench
=====================================================================

lmbench outputs status information as it runs various benchmarks.
By default this output is sent to /dev/tty, but you may redirect
it to any file you wish (such as /dev/null...).

Status output file [default /dev/tty]: 
=====================================================================

There is a database of benchmark results that is shipped with new
releases of lmbench.  Your results can be included in the database
if you wish.  The more results the better, especially if they include
remote networking.  If your results are interesting, i.e., for a new
fast box, they may be made available on the lmbench web page, which is

        http://www.bitmover.com/lmbench

Mail results [default yes]: no
OK, no results mailed.
=====================================================================

Confguration done, thanks.

There is a mailing list for discussing lmbench hosted at BitMover. 
Send mail to majordomo@bitmover.com to join the list.

Using config in CONFIG.sbcx
Thu Jan 24 09:27:32 CET 2020
Latency measurements
Thu Jan 24 09:28:38 CET 2020
Local networking
Fri Jan 24 09:29:38 CET 2020
Bandwidth measurements
Fri Jan 24 09:41:01 CET 2020
Calculating context switch overhead
Fri Jan 24 09:41:06 CET 2020
Calculating effective TLB size
Fri Jan 24 09:41:08 CET 2020
Calculating memory load parallelism
Fri Jan 24 09:53:02 CET 2020
McCalpin's STREAM benchmark
Fri Jan 24 09:54:55 CET 2020
Calculating memory load latency
Fri Jan 24 10:28:12 CET 2020
Benchmark run finished....
Remember you can find the results of the benchmark 
under /var/lib/lmbench/results

Results

TBD inserire link a file scaricabile