| View previous topic :: View next topic |
| Author |
Message |
Sakaki
Guru
Joined: 21 May 2014
Posts: 409
|
 Posted: Fri Jul 06, 2018 12:29 pm Post subject: |
 |
|
| orion777 wrote: | ...
Now I was copying my custom kernel to the /boot
but I forgot to run make modules_install again!
...
I'm still able to access the file system via card reader of the laptop, but I don't know how to solve this problem  If someone have any ideas, these are welcome, because reinstalling again and again.. rpi3 64 bit image, kernel implementation and custom drone software are time-expensive.. |
Probably the simplest thing, to get going again, would be to mount your microSD card's partition 1 and 2 on your Linux PC, download the current prebuilt bcmrpi3-kernel-bis tarball, and untar it onto the card (per the instructions here). This should be enough to get you booted, you can then install the modules again. None of this requires any cross-compiling etc.
For example, assuming your card shows up as (e.g.) /dev/mmcblk0 on your your PC, you could do (adapt as appropriate):
Mount the card partitions on your PC: | Code: | gentoo-pc ~ # mkdir -pv /mnt/piroot
gentoo-pc ~ # mount -v /dev/mmcblk0p2 /mnt/piroot
gentoo-pc ~ # mount -v /dev/mmcblk0p1 /mnt/piroot/boot |
Archive your custom kernel, download latest binary kernel tarball (it is <15MiB):
| Code: | gentoo-pc ~ # mv /mnt/piroot/boot/kernel8.img{,.old}
gentoo-pc ~ # wget -c https://github.com/sakaki-/bcmrpi3-kernel-bis/releases/download/4.14.52.20180703/bcmrpi3-kernel-bis-4.14.52.20180703.tar.xz |
Untar the kernel and module set (take care the path after the -C is correct!), ensure written, and dismount:
| Code: | gentoo-pc ~ # tar -xJf bcmrpi3-kernel-bis-4.14.52.20180703.tar.xz -C /mnt/piroot/
gentoo-pc ~ # sync
gentoo-pc ~ # umount -v /mnt/piroot/boot
gentoo-pc ~ # umount -v /mnt/piroot |
Once rebooted, you can move /boot/kernel8.img.old back to /boot/kernel8.img, do a make modules_install in your kernel build directory, and reboot.
You should then have a functioning system.
| orion777 wrote: | | Is it possible to exclude modemmanager from the system? |
Then, to remove modemmanager, the simplest thing is just to add USE="-modemmanager" to /etc/portage/make.conf (if you already have any USE flags set in there, simply append -modemmanager to the list of course, prefixed with a space), and issue:
| Code: | pi64 ~ # emerge --verbose --deep --with-bdeps=y --update --changed-use @world
pi64 ~ # emerge --depclean |
And it should be gone; no need for masking etc.
PS to find out what is pulling in a package, you can use the bundled equery tool, for example:
| Code: | pi64 ~ # equery depends net-misc/modemmanager
* These packages depend on net-misc/modemmanager:
gnome-extra/nm-applet-1.8.10-r1 (modemmanager ? net-misc/modemmanager)
net-misc/networkmanager-1.10.10 (modemmanager ? >=net-misc/modemmanager-0.7.991:0) |
As you can see there are two packages with direct deps, but these are conditional on the modemmanager USE flag (per the <useflag> ? <dep> syntax). As such, turning this flag off (as above) and rebuilding affected packages (the emerge ... --changed-use @world) will remove any dependencies on the installed net-misc/modemmanager instance, and emerge --depclean will then purge it from your system.
_________________
Regards,
sakaki |
|
| Back to top |
|
 |
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Sat Jul 07, 2018 7:19 am Post subject: |
|
|
| Quote: | Untar the kernel and module set (take care the path after the -C is correct!), ensure written, and dismount:
gentoo-pc ~ # tar -xJf bcmrpi3-kernel-bis-4.14.52.20180703.tar.xz -C /mnt/piroot/ |
Where the kernel8.img must be located? In the boot partition or in the root partition folder boot?
Probably in the boot partition, as welcome screen kernel version changes only if I put the kernel into the boot partition.
Whatever, nor boot partition nor root partition boot folder does not solve my problem..
Maybe there is a hint that the system become (none) instead of pi64? Where system naming is done? maybe I will find some missing files there..
| Quote: | | to remove modemmanager, the simplest thing is just to add USE="-modemmanager" |
Sorry, I forgot about USE flags, that can be used in gentoo  |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Sat Jul 07, 2018 7:22 am Post subject: |
|
|
| Goverp wrote: |
While I've not tried on this system (only just installed it and hardly played with it yet), my (AMD) laptop is KDE-based, and uses wpa_supplication and its GUI alone to manage WiFi. Neither networkmanager nor modemmanager are installed, nor whatever else they drag in. |
As I know, it is possible to run only dhcpcd and it will manage all interfaces alone (excluding WiFi network selection). |
|
| Back to top |
|
|
NeddySeagoon
Administrator
Joined: 05 Jul 2003
Posts: 54588
Location: 56N 3W
|
| Posted: Sat Jul 07, 2018 8:56 am Post subject: |
|
|
orion777,
The boot process on the Pi is as follws. The VC4 video chip runs its internal fimware, which you cannot acoess, to load bootcode.bin into RAM from the first partition on the SD card which must be vfat. The ARM CPU is held reset.
Control is passed to bootcode.bin which reads config.txt to learn how to adjust its built in defaults.
Once the kernel, initrd (if any), and dtb are all loaded the ARM CPU is released from reset.
Getting started is all down to the VC4 and bootcode.bin. The ARM CPU is kept out of the way until the hard bit is done.
bootcode.bin looks at the hardware it finds itself on and adjusts the defaults to suit.
_________________
Regards,
NeddySeagoon
Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail. |
|
| Back to top |
|
|
Sakaki
Guru
Joined: 21 May 2014
Posts: 409
|
| Posted: Sat Jul 07, 2018 10:17 am Post subject: |
|
|
| orion777 wrote: | Where the kernel8.img must be located? In the boot partition or in the root partition folder boot?
Probably in the boot partition, as welcome screen kernel version changes only if I put the kernel into the boot partition.
Whatever, nor boot partition nor root partition boot folder does not solve my problem..
Maybe there is a hint that the system become (none) instead of pi64? Where system naming is done? maybe I will find some missing files there.. |
Maybe I have mis-understood your issue, if so apologies.
To boot a kernel successfully, you need various firmware files on the microSD card's first partition's vfat filesystem (bootcode.bin etc.), the kernel itself (kernel8.img) and its dtbs (same filesystem), plus various config files (cmdline.txt etc) (same filesystem). Let's call this filesystem bootfs. You also need, on the microSD card's second partition's filesystem (ext4 as shipped, let's call this rootfs), under /lib/modules, a matching module set for the kernel in question.
Once the gentoo-on-rpi3-64bit image is booted, the rootfs is mounted at /, and the bootfs at /boot within it.
Since you wrote:
| orion777 wrote: | Now I was copying my custom kernel to the /boot
but I forgot to run make modules_install again!
...
I'm able to login, but there is no network interfaces, none of commands like make menuconfig, make modules_install are working, returning an error read-only file system
...
Original kernel probably boot the system (as it responds on ctrl+alt+del), but the screen is empty (probably video chip is not started) |
I provided you instructions to manually install a working kernel + module set from a tarball onto the microSD card using a PC.
The tarball paths by default assume that the rootfs is mounted at ./ and the bootfs is mounted at ./boot (within ./).
However, on the PC following the above instructions, the rootfs will be mounted at /mnt/piroot and the bootfs at /mnt/piroot/boot
During the untar step, the working directory is changed (via -C) to /mnt/piroot/, ensuring that the files in the archive get put in the correct place (the kernel and dtbs into the bootfs, and the matching module set into /lib/modules within the rootfs).
The point of this is to get back to a kernel + modules system that will boot without errors, without losing the any other work stored on the rootfs, requiring any cross-compilation, or download of large files (the tarball mentioned is < 15MiB).
Note that during boot, by default, the rootfs is initially mounted r/o and is later changed to r/w by an OpenRC boot service (if the initial boot has gone OK), /etc/init.d/root. Missing modules will likely cause this to fail, leaving you with a r/o filesystem.
Some other miscellaneous troubleshooting hints:
If you can get a login console on your rpi3 after booting, you can log in as root and try issuing:
| Code: | # mount -n -o remount,rw /
# mount /boot |
To manually remount the rootfs r/w, then mount the bootfs at /boot within it. This may be enough to allow you to make modules_install or whatever and then reboot.
Also, if your screen is blank, try commenting out the following line in config.txt on the bootfs, so it reads:
| Code: | | #dtoverlay=vc4-fkms-v3d,cma-256 |
Leave the rest of the file as-is. You can do this with the bootfs mounted on a PC etc. This will use the simpler graphics driver on boot so you should see error messages etc. earlier (you need the vc4 graphics driver module otherwise).
The hostname is set by the OpenRC service /etc/init.d/hostname, incidentally.
_________________
Regards,
sakaki |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Mon Jul 09, 2018 6:13 pm Post subject: |
|
|
Thank You for the explanation!
Ohh, now I understand, this was confusing me, since I was first time worked from the card reader..
| Sakaki wrote: |
Once the gentoo-on-rpi3-64bit image is booted, the rootfs is mounted at /, and the bootfs at /boot within it.
|
I will recheck all files. Also I have another flash card with your image, so I should to be able to compare. This flash was prepared a few days ago; also genup was runned and the system is still able to operate correctly.
| Quote: | | Note that during boot, by default, the rootfs is initially mounted r/o and is later changed to r/w by an OpenRC boot service (if the initial boot has gone OK), /etc/init.d/root. Missing modules will likely cause this to fail, leaving you with a r/o filesystem. |
Yes, it was fond by me into the internet; I will check this service too. Thank You. |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Fri Jul 13, 2018 4:46 pm Post subject: |
|
|
dev-python/matplotlib-2.2.2-r1 can be installed (at least to me) only with x11-libs/wxGTK 3.0.3(3.0){tbz2}(20:26:31 15/01/18)(X libnotify opengl tiff -aqua -debug -doc -gstreamer -sdl ABI_MIPS="-n32 -n64 -o32" ABI_PPC="-32 -64" ABI_S390="-32 -64" ABI_X86="-32 -64 -x32")
If x11-libs/wxGTK 3.0.4 is installed, then matplotlib-2.2.2-r1 compillator will return
| Code: | Fatal Error: Mismatch between the program and library build versions detected.
The library used 3.0 (wchar_t,compiler with C++ ABI 1010,wx containers,compatible with 2.8),
and wxPython used 3.0 (wchar_t,compiler with C++ ABI 1011,wx containers,compatible with 2.8). |
|
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
|
| Back to top |
|
|
Sakaki
Guru
Joined: 21 May 2014
Posts: 409
|
| Posted: Sun Jul 15, 2018 10:47 am Post subject: |
|
|
orion777,
on the image, /usr/local/portage is used as a top-level directory to hold multiple local ebuild repositories (aka overlays), so please don't create an ebuild tree at the top level in there. Instead, you just need to follow the instructions you linked above to create "localrepo", but change the path to use a subdirectory (e.g. to /usr/local/portage/localrepo or similar; make sure to reflect this path in /etc/portage/repos.conf/localrepo.conf, and you'd do e.g. "mkdir -p /usr/local/portage/localrepo/{metadata,profiles}" etc). You can use thin manifests in your own repo to avoid the need to create manifest entries for the ebuilds themselves (you do this in /usr/local/portage/localrepo/metadata/layout.conf - see /usr/local/portage/sakaki-tools/metadata/layout.conf for an example).
_________________
Regards,
sakaki |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Tue Jul 17, 2018 7:40 am Post subject: |
|
|
I don't know what was happening here, but I'm still get an error :/
| Code: |
mkdir -p /usr/local/portage/localrepo/{metadata,profiles}
chown -R portage:portage /usr/local/portage/localrepo
cat /usr/local/portage/localrepo/profiles/repo_name
localrepo
cat /usr/local/portage/localrepo/metadata/layout.conf
repo-rame = localrepo
masters = gentoo
auto-sync = false
cat /etc/portage/repos.conf/localrepo.conf
[localrepo]
location = /usr/local/portage/localrepo
mkdir -p /usr/local/portage/localrepo/ros-meta/mavproxy
cp ~root/mavproxy-9999.ebuild /usr/local/portage/localrepo/ros-meta/mavproxy/mavproxy-9999.ebuild
chown -R portage:portage /usr/local/portage/localrepo ## Again? But Ok..
pushd /usr/local/portage/localrepo/ros-meta/mavproxy
pi64 /usr/local/portage/localrepo/ros-meta/mavproxy # repoman manifest
* ERROR: ros-meta/mavproxy-9999::localrepo failed (depend phase):
* External commands disallowed while sourcing ebuild:
*
* Call stack:
* ebuild.sh, line 635: Called source '/usr/local/portage/localrepo/ros-meta/mavproxy/mavproxy-9999.ebuild'
'* mavproxy-9999.ebuild, line 3: Called command_not_found_handle '
* ebuild.sh, line 88: Called die
* The specific snippet of code:
* die "External commands disallowed while sourcing ebuild: ${*}"
*
* If you need support, post the output of `emerge --info '=ros-meta/mavproxy-9999::localrepo'`,
* the complete build log and the output of `emerge -pqv '=ros-meta/mavproxy-9999::localrepo'`.
* Working directory: '/usr/lib64/python3.5/site-packages'
* S: '/var/tmp/portage/ros-meta/mavproxy-9999/work/mavproxy-9999'
!!! getFetchMap(): aux_get() error reading ros-meta/mavproxy-9999; aborting.
Unable to generate manifest.
|
| Code: | pi64 /usr/local/portage/localrepo/ros-meta/mavproxy # cat /usr/local/portage/localrepo/ros-meta/mavproxy/mavproxy-9999.ebuild
# Copyright 1999-2017 Gentoo Foundation
# Distributed under the terms of the GNU General Public License v2
EAPI=6
PYTHON_COMPAT=( python2_7 )
inherit distutils-r1 git-r3
DESCRIPTION="A UAV ground station software package for MAVLink based systems"
HOMEPAGE="http://ardupilot.github.io/MAVProxy"
EGIT_REPO_URI="https://github.com/ArduPilot/MAVProxy.git"
LICENSE="GPL-3+"
SLOT="0"
# KEYWORDS="~amd64 ~x86"
REQUIRED_USE="${PYTHON_REQUIRED_USE}"
RDEPEND="${PYTHON_DEPS}
media-libs/opencv[${PYTHON_USEDEP}]
dev-python/wxpython:2.8[${PYTHON_USEDEP}]
dev-python/matplotlib[${PYTHON_USEDEP}]
dev-python/numpy[${PYTHON_USEDEP}]
dev-python/pyserial[${PYTHON_USEDEP}]
dev-python/pillow[${PYTHON_USEDEP}]
"
DEPEND="${RDEPEND}
virtual/pkgconfig" |
|
|
| Back to top |
|
|
Sakaki
Guru
Joined: 21 May 2014
Posts: 409
|
| Posted: Tue Jul 17, 2018 4:21 pm Post subject: |
|
|
orion777,
I just tried setting up 'localrepo' (containing mavproxy-9999.ebuild) on a fresh copy of the image, following as exactly as I could the steps you describe above. After installing app-portage/repoman, I was able to run "repoman manifest" without error (an alternative, incidentally, would be to run "ebuild mavproxy-9999.ebuild manifest").
Perhaps you have altered some other part of the config of your system that is leading to this issue?
_________________
Regards,
sakaki |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Tue Jul 17, 2018 7:47 pm Post subject: |
|
|
Okay.. I have this error with different ebuild and its folder names..
So.. how did You create the mavproxy-9999.ebuild file in your system?
-------------
UPDATE - It is fixed!
So: I was copying ebuild text from https://forums.gentoo.org/viewtopic-p-8241910.html#8241910 and paste it to the windows notepad; then it was uploaded to the linux and...
cat shows that everything is good; mc - view (F3) also shows that everything is perfect; but if you will open the ebuild file (created in windows) via mc - edit (F4), then you will see this:
https://ibb.co/hvDBmy Just clean this unnecessary stuff and feel lucky )) |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Sat Jul 21, 2018 8:47 am Post subject: |
|
|
| Wish to say Thanks to Sakaki for the help in ebuild testing! Everything is working. |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Tue Jul 24, 2018 7:35 am Post subject: |
|
|
| Good day! Is it possible to install net-wireless/hostapd ? The problem is that it forses -bindist flag for openssl, dev-embedded/rpi3-64bit-meta-1.2.2-r6 and so... |
|
| Back to top |
|
|
roylongbottom
n00b
Joined: 13 Feb 2017
Posts: 64
Location: Essex, UK
|
| Posted: Thu Jul 26, 2018 10:45 am Post subject: Raspberry Pi 3B+ Multithreading Benchmarks |
|
|
Raspberry Pi 3B+ Multithreading Benchmarks
This is a summary of 32 bit and 64 bit results on Rpi 3B and 3B+, run via Raspbian Stretch and the latest version of 64 bit Gentoo. Full details are available at ResearchGate in:
https://www.researchgate.net/publication/324694380_Raspberry_Pi_3B_32_Bit_and_64_Bit_Benchmarks_and_Stress_Tests
Except for the OpenMP benchmarks, all run using 1, 2, 4 and 8 threads, executing common code. One [/b]and four thread results are provided here.
MP-Whetstone Benchmark
This has an overall rating in MWIPS, mainly dependent on floating point calculations shown under MFLOPS, Cos MOPS and Exp MOPs, with the latter MOPS currently having the most impact. Each thread used dedicated variables to provide best case multithreading speeds.
Based on MWIPS, in all cases, 4 thread performance was around four times single thread performance and 3B+ speed about 17% faster than that for 3B, proportional to CPU MHz ratio (1400/1200). Test functions are too simple to be helped much by advanced architecture, leading to 64 bit benchmark average performance being 6% faster than the 32 bit compilation.
| Code: | MWIPS MFLOPS MFLOPS MFLOPS Cos Exp Fixpt If Equal
1 2 3 MOPS MOPS MOPS MOPS MOPS
32 Bit
3B 1T 924.2 335.9 276.8 298.5 18.5 10.4 5817.2 1035.3 719.4
4T 3718.4 1286.3 1303.9 1193.5 74.3 41.5 19961.4 4698.4 2862.7
3B+ 1T 1084.2 391.0 384.9 348.6 21.7 12.1 6967.0 1013.1 822.3
4T 4343.8 1540.9 1558.3 1389.5 86.6 48.4 27529.5 5549.8 3338.5
64 Bit
3B 1T 979.8 330.4 322.9 281.5 20.0 10.8 1368033.3 2335.5 1177.1
4T 3914.5 1206.2 1295.8 1122.2 78.3 44.3 3007162.3 9230.2 4636.6
#########
3B+ 1T 1151.6 383.0 382.7 327.6 23.2 13.0 1717931.5 2720.5 1364.5
4T 4579.6 1505.5 1525.7 1304.4 92.0 51.6 4647842.5 10777.1 5448.5
######### over optimised but little effect on MWIPS rating
|
MP-Dhrystone Benchmark
This runs multiple copies of the whole program at the same time. Dedicated data arrays are used for each thread but there are numerous other variables that are shared. The latter can reduce performance gains via multiple threads and, in some cases, these can be slower than using a single thread (not in this case).
Comparisons are shown below, as expected 3B+/3B, MP performance not good, 64 bit faster than 32 bit - additional registers provide more options for optimisation (or over optimisation for which this benchmark is notorious).
| Code: | Rating in VAX MIPS aka DMIPS
Threads 1 2 4 8 4T/1T
32 Bit
3B 2338 3956 5730 5780 2.45
3B+ 2694 4720 6716 6729 2.49
3B+/3B 1.15 1.19 1.17 1.16
64 Bit
3B 3622 5752 6290 6359 1.74
3B+ 4198 6158 7354 7411 1.75
3B+/3B 1.16 1.07 1.17 1.17
64/32 bit
3B 1.55 1.45 1.10 1.10
3B+ 1.56 1.30 1.09 1.10
|
MP-Linpack Benchmark
The original Linpack benchmark, operates on double precision floating point 100x100 matrices (N = 100). This version uses NEON intrinsic functions with single precision calculations. It is run run on 100x100, 500x500 and 1000x1000 matrices and includes a test without threading overheads. The benchmark was produced to demonstrate that the original Linpack 100x100 code could not be converted (by me) to show increased performance using multiple threads. The official line is that users are allowed to use their own linear equation solver for this purpose. These Raspbian tests were carried out under the later Stretch release.
Performance can vary somewhat with this benchmark but reflects the usual 3B+ speed gains, at least on averaging all results. On the same basis, average 64 bit speeds are suggested as being the same as those at 32 bits, but some indicate slower performance. Similar performance could be expected as the compiled code is derived from high level NEON SIMD vector functions.
The poor performance, even using a single thread, is due to the frequent starting and stopping of threads to execute the critical calculations. Consistent threaded speed indicates shared data write back to RAM dependency. This probably increases with larger matrices as more calculations are carried out during a threaded function call.
| Code: | Threads None 1 2 4
32 bit 3B
N 100 542.22 61.00 60.67 60.74
N 500 480.55 311.06 316.00 303.48
N 1000 364.07 272.49 231.10 232.07
32 bit 3B+
N 100 633.11 70.82 70.13 70.20
N 500 505.37 323.24 326.81 327.73
N 1000 378.29 337.34 337.01 337.80
64 bit 3B
N 100 551.48 87.43 81.66 82.68
N 500 359.51 258.43 242.92 255.61
N 1000 296.11 281.75 279.20 282.71
64 bit 3B+
N 100 639.82 100.30 95.24 95.25
N 500 430.41 292.80 291.12 290.04
N 1000 349.47 313.59 312.38 313.40
|
MP-BusSpd Benchmark
This runs integer read only tests using caches and RAM, each thread accessing the same data sequentially, in this case, each thread starts reading from different addresses. To start with, data is read with large address increments to demonstrate burst data transfers. Below are results for reading all data.
Most striking is the poor relative performance of the 64 bit version, with data from L1 caches. this is due to the compile options used, with the 32 bit program using four way multiple data vector instructions, but with the 64 bit compiler producing single data scalar. Using the later gcc 7 compiler made no difference.
The usual 3B+/3B gains can be assumed to apply, along with 4 core / 1 core throughput increases from cache based data, with those from RAM indicting some improvement.
| Code: | MB/Second Reading All Data, 1, 2, 4, 8 Threads
KB and 32 Bit 64 Bit
Theads 3B 3B+ 3B 3B+
12.3 1T 4223 4928 1492 1731
2T 8159 9648 2928 3404
4T 15720 18252 5346 5524
8T 13789 16504 5016 5780
122.9 1T 3934 4565 1470 1694
2T 7686 8962 2932 3408
4T 13860 15935 5261 6781
8T 13125 15826 5271 6250
12288 1T 1906 2085 1024 1048
2T 1978 2126 1590 2025
4T 2135 2338 1546 1623
8T 2158 2488 1821 2018
|
MP-RandMem Benchmark
The benchmark has cache and RAM read only and read/write tests using sequential and random access, each thread accessing the same data but starting at different points. Although performance via the L1 cache, L2 cache and RAM can be different, it is normally consistent, in each of these areas, during read/write tests.
On all tests, 3B+/3B performance gains were as expected for cache based results, with averages between 1.6 and 1.7, with RAM performance being similar. Read only MP gains were mainly greater than 3.5 times for cache tests, except on random access to L2, at around 2.4 times, understandably lower using a shared cache. There were also some MP increased throughput using RAM. Raspbian based results indicate slightly improved performance over those using 64 bit Gentoo.
| Code: |
MB/Second Using 1, 2, 4 and 8 Threads
3B 32 Bit 3B+ 32 Bit
KB SerRD SerRW RndRD RndRW SerRD SerRW RndRD RndRW
12.3 1T 4078 3814 4018 3798 4747 4447 4776 4435
2T 8045 3768 8043 3777 9253 4362 9378 4362
4T 15622 3724 15625 3730 18114 4343 18080 4322
8T 15208 3723 15020 3724 17813 4345 17788 4321
122.9 1T 3289 3393 827 891 3871 3893 948 1016
2T 6556 3379 1512 880 7612 3954 1742 1021
4T 12125 3364 2078 886 14399 3929 2383 1025
8T 12309 3364 2042 886 14089 3935 2367 1023
12288 1T 1669 878 65 64 1850 860 67 68
2T 3485 872 121 65 3670 867 126 67
4T 4296 876 146 65 4097 874 146 68
8T 2435 878 147 65 2919 873 148 68
3B 64 Bit 3B+ 64 Bit
12.3 1T 4260 3071 4261 3081 4939 3573 4941 3574
2T 7500 3054 7496 3059 8730 3553 8704 3545
4T 15092 3018 14794 3019 17121 3499 17197 3498
8T 14315 2977 14544 2989 16685 3454 17097 3471
122.9 1T 3385 2861 867 837 3936 3347 1014 975
2T 6323 2653 1543 838 7334 3344 1794 976
4T 10638 2873 2009 835 12475 3333 2281 973
8T 10810 2841 1947 834 12261 3314 2323 974
12288 1T 1607 746 71 60 1921 793 77 63
2T 1605 696 123 59 1896 784 113 58
4T 1939 766 129 58 1621 707 126 58
8T 1682 681 141 58 1302 764 106 57
|
MP-MFLOPS Benchmarks
The benchmarks carry out calculations using 2 and 32 floating point operations per data word read. The latter is of the form form x[i] = (x[i] + a) * b - (x[i] + c) * d + (x[i] + e) * f + ............), to take advantage of linked multiply and add instructions that, with SIMD, could produce 32 single precision MFLOPS per MHz (4 words at a time x 4 cores x 2 for linked instructions). The same benchmark, compiled with gcc Linux, obtains 23 MFLOPS/MHz on an Intel Core i7 four core processor, using SSE instructions, but it is not clear what theoretical sustained rate that could be obtained on a Raspberry Pi 3.
There are three each of 32 bit and 64 bit programs provided, Single Precision (SP), Double Precision (SP) and one using NEON Intrinsic Functions (SP). A fourth variation was produced using OpenMP, that also includes tests using 8 operations per word. Each thread uses the same calculations but accessing different segments of the data. The program checks for consistent numeric results, primarily to show that all calculations are carried out and can be run. The standard format is shown below.
| Code: |
MP-MFLOPS Gentoo RPi 3B+ 64 Bit
MP-MFLOPS armv8 64Bit Wed Apr 25 10:22:43 2018
FPU Add & Multiply using 1, 2, 4 and 8 Threads
2 Ops/Word 32 Ops/Word
KB 12.8 128 12800 12.8 128 12800
MFLOPS
1T 824 763 380 1793 1795 1704
2T 1620 1577 404 3584 3588 3324
4T 2093 2933 387 6981 6843 4189
8T 2481 2263 409 6851 6391 2944
Results x 100000
1T 76406 97075 99969 66015 95363 99951
2T 76406 97075 99969 66015 95363 99951
4T 76406 97075 99969 66015 95363 99951
8T 76406 97075 99969 66015 95363 99951
End of test Wed Apr 25 10:22:48 2018
|
Calculations show that the usual 3B+ performance gains were produced on all benchmarks and that maximum performace was always using 32 operations per word. Indications of maximum performance from the various benchmarks is provided below, where most MP gains are reasonable. Best case was using NEON Intrinsic Functions at 8.26 MFLOPS/MHz (a long way from 32). Note the 64/32 bit speed gains with the original benchmarks. These are typical for all tests, excluding those wit RAM based data.
The original 32 bit versions converted the 32 operations into 22 scalar instructions, with 10 fused multiply accumulate type. The 64 bit variations unrolled the execution loop to include 128 operations, comprising 22 four way vector instructions, but speed of the original versions were degraded through using too few registers with too many register loads. A new 64 bit program is now provided, compiled with gcc 7, and this provides the same order of performance as the NEON benchmark, using 25 vector registers (4 way SP) with out of loop loads.
| Code: | Raspberry Pi 3B+ MFLOPS at 32 Operations Per Data Word
NEON 64 bit OpenMP
32 bit 64 bit 64 bit 64 bit gcc7 gcc6 gcc7
SP DP SP DP SP SP DP SP SP
1 Thread 813 798 1793 1405 2999 2800 1403 1692 2781-@
4 Threads 3189 3109 6981 4398 11563 10608 4492 6469 10007
4T/1T 3.92 3.90 3.89 3.13 3.86 3.79 3.20 3.82 3.60
|
OpenMP Benchmarks
These benchmarks use OpenMP complier directives and programming options to automatically spread the load over multiple CPUs or Cores. Two benchmarks are provided, one named notOpenMP-name, that is compiled from the same code without using OpenMP options, to provide single core performance.
OpenMP-MFLOPS Benchmark
This benchmark carries out the same calculations as the MP-MFLOPS Benchmarks but, in addition, calculations with eight operations per data word. As indicated in the above table, recompiling with gcc 7 produced similar performance as the best C version. Note, these benchmarks are from the same code as the Linux PC versions, with higher memory demands than MP-MFLOPS, starting at 400 KB, L2 cache size. Below are single and multi-core 64 bit 3B+ results, indicating that MP performance was much slower when RAM data was involved.
Comparisons are also provided, indicating the usual 3B+/3B performance gains, at least on using a single core. Respectable MP gains were obtained at 64 bits but less so at 32 bits. The latter leads to indications that 64 bit working is faster on multi-tasking, but the detail shows that 32 bit speeds are faster using a single core.
| Code: |
notOpenMP MFLOPS64 GCC7 Mon Jul 9 11:28:16 2018
Test 4 Byte Ops/ Repeat Seconds MFLOPS First All
Words Word Passes Results Same
Data in & out 100000 2 2500 0.646220 774 0.929538 Yes
Data in & out 1000000 2 250 1.199940 417 0.992550 Yes
Data in & out 10000000 2 25 1.158499 432 0.999250 Yes
Data in & out 100000 8 2500 1.049060 1906 0.957117 Yes
Data in & out 1000000 8 250 1.403440 1425 0.995518 Yes
Data in & out 10000000 8 25 1.375663 1454 0.999549 Yes
Data in & out 100000 32 2500 2.876464 2781 0.890215 Yes
Data in & out 1000000 32 250 3.176167 2519 0.988088 Yes
Data in & out 10000000 32 25 3.171387 2523 0.998796 Yes
End of test Mon Jul 9 11:28:33 2018
OpenMP MFLOPS64 GCC7 Mon Jul 9 11:25:46 2018
Data in & out 100000 2 2500 0.250953 1992 0.929538 Yes
Data in & out 1000000 2 250 1.185075 422 0.992550 Yes
Data in & out 10000000 2 25 1.194995 418 0.999250 Yes
Data in & out 100000 8 2500 0.332016 6024 0.957117 Yes
Data in & out 1000000 8 250 1.182464 1691 0.995518 Yes
Data in & out 10000000 8 25 1.157733 1728 0.999549 Yes
Data in & out 100000 32 2500 0.799423 10007 0.890215 Yes
Data in & out 1000000 32 250 1.260969 6344 0.988088 Yes
Data in & out 10000000 32 25 1.196507 6686 0.998796 Yes
End of test Mon Jul 9 11:25:55 2018
Comparisons
Raspbian 32 Bit Gentoo 64 Bit Gentoo 64 Bit gcc 7
3B 3B+ 3B+/3B 3B 3B+ 3B+/3B 3B 3B+ 3B+/3B
8 Ops/Wd
1 Core 1723 2005 1.16 1623 1892 1.17 1654 1906 1.15
4 Cores 4485 5071 1.13 5713 6504 1.14 5893 6024 1.02
4/1 Cores 2.60 2.53 3.52 3.44 3.56 3.16
32 Ops/Wd
1 Core 1592 1857 1.17 1461 1692 1.16 2395 2781 1.16
4 Cores 1858 2420 1.30 5493 6469 1.18 8722 10007 1.15
4/1 Cores 1.17 1.30 3.76 3.82 3.64 3.60
|
OpenMP-MemSpeed Benchmark
This is the same as the single core Memory Speed Benchmark, with similar results to NotOpenMP varieties, but with measurements extending to test more memory, with 16 sets of results between 4 KB and 128 MB. Floating point and integer calculations are carried out as indicated in the table headings. Samples of results are provided below, covering L1 cache, L2 cache and RAM based data.
Comparing characteristics provides wide ranges of variations. The 3B+/3B ratio was almost entirely in the expected 1.6 to 1.7 ratio for L1 cache data, but with some significant differences via L2 cache and RAM. On comparing MP performance, the most noticeable were many at half speed with integer arithmetic. Then, 32 bit Raspbian produced more than three times increase on double precision calculations and twice at single precision. The 64 bit Gentoo floating MP gains were around 1.5 times. These strange results were highlighted in 64/32 bit comparisons, where the former tended to be faster standalone but slower in MP mode. Alternative compilations might be of interest.
| Code: | Memory x[m]=x[m]+s*y[m] Int+ x[m]=x[m]+y[m] x[m]=y[m]
KBytes Dble Sngl Int32 Dble Sngl Int32 Dble Sngl Int32
Used MB/S MB/S MB/S MB/S MB/S MB/S MB/S MB/S MB/S
Raspbian 32 Bit
Pi 3B Not
8 1594 2547 3811 2388 3469 3812 2717 2716 2716
16 1595 2553 3825 2393 3478 3825 2728 2728 2728
64 1508 2300 3304 2177 3065 3303 2542 2485 2485
128 1515 2305 3353 2183 3108 3356 2644 2573 2574
32768 928 1166 1206 1119 1224 1206 760 746 746
65536 970 1171 1210 1140 1225 1212 811 810 808
Pi 3B OMP
8 6068 3107 1382 10109 5056 1486 16438 8104 1258
16 5739 3119 1317 10193 5114 1393 16624 7862 1220
64 5416 3055 1303 8618 4928 1403 12254 8045 1218
128 5396 3050 1359 9101 4932 1379 9496 8089 1249
32768 3351 2499 1080 2089 3216 1437 1005 1001 794
65536 3820 614 1026 3901 3078 1209 1006 1008 954
Pi 3B+ Not
8 1860 2973 4449 2787 4047 4447 3169 3169 3170
16 1862 2978 4463 2791 4058 4462 3137 3183 3182
64 1753 2664 3806 2526 3541 3805 3016 3028 3046
128 1776 2721 3968 2574 3683 3965 3073 2989 2989
32768 939 1218 1205 1192 1219 1213 909 864 864
65536 1052 1221 1216 1197 1214 1213 905 870 869
Pi 3B+ OMP
8 6142 3447 1613 11721 5857 1721 19080 9581 1465
16 6315 3608 1568 11687 5911 1699 19295 9528 1468
64 5795 3522 1456 10640 5718 1627 14169 9392 1397
128 5657 3542 1485 10485 5713 1588 11330 9181 1398
32768 1388 2107 1090 4044 3099 1355 941 933 1358
65536 2576 1750 1159 2296 3825 1661 943 971 1509
Gentoo 64 Bit
Pi 3B Not
8 3935 2507 4155 5339 3399 4168 4646 3730 3732
16 3857 2506 4171 5358 3417 4155 4680 3744 3751
64 3074 2242 3416 3956 2940 3438 4432 3635 3618
128 3042 2274 3495 3933 3008 3492 4209 3488 3498
32768 529 1059 939 1106 975 898 695 701 495
65536 1095 986 1017 990 1104 694 682 686 625
Pi 3B OMP
8 6064 3205 1690 9310 4858 1824 10003 5704 2052
16 3584 3228 1634 8855 4902 1742 10164 5598 1856
64 3697 2665 1563 8723 4798 1655 10674 5331 1758
128 2947 3206 1672 8900 4841 1618 9060 5456 1768
32768 605 792 1562 1143 2709 1019 899 861 591
65536 1312 555 1163 3974 803 1181 908 833 780
Pi 3B+ Not
8 4625 2906 4828 6167 3957 4828 5384 4307 4318
16 4624 2902 4840 6203 3954 4838 5422 4347 4332
64 3857 2704 4232 4873 3589 4202 5002 4121 4122
128 3899 2760 4371 4906 3720 4373 5030 4079 4111
32768 498 1107 1064 1197 1202 1135 799 649 763
65536 1176 1060 1165 1177 1203 1230 588 959 963
Pi 3B+ OMP
8 7385 3728 1976 8941 5652 2107 13290 6637 2392
16 7099 3769 1945 11092 5751 2105 11593 2798 993
64 1432 1264 801 4343 2345 852 3740 2531 914
128 4476 3711 1812 9677 5654 2070 4091 2665 846
32768 777 1994 1471 1223 1269 1800 958 945 790
65536 1736 1275 1271 1006 1249 1787 869 980 1293
|
_________________
Regards
Roy |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Sun Jul 29, 2018 8:02 am Post subject: |
|
|
| orion777 wrote: | | Good day! Is it possible to install net-wireless/hostapd ? The problem is that it forses -bindist flag for openssl, dev-embedded/rpi3-64bit-meta-1.2.2-r6 and so... |
So, I was creating an overlay based on hostapd version 2.6-r6 and call it 2.6-r10. In this overlay I delete [-bindist] requirement from openssl At least for now it is working, but will see.. |
|
| Back to top |
|
|
orion777
Apprentice
Joined: 15 Mar 2017
Posts: 207
Location: Riga, Latvia
|
| Posted: Sun Jul 29, 2018 8:06 am Post subject: |
|
|
| Maybe someone know, how to send shutdown command to the rpi3 if there are no ssh and attached usb keyboard? Because I have one raspberry as a NAT server, like "plug and it will boot", but how to stop them? |
|
| Back to top |
|
|
Sakaki
Guru
Joined: 21 May 2014
Posts: 409
|
| Posted: Mon Jul 30, 2018 6:56 pm Post subject: |
|
|
All,
Apologies, I am currently travelling with only very intermittent internet access; I'll be able to respond to support questions again upon my return (last week of August).
_________________
Regards,
sakaki |
|
| Back to top |
|
|
Goverp
Advocate
Joined: 07 Mar 2007
Posts: 2182
|
| Posted: Sun Aug 12, 2018 7:31 pm Post subject: |
|
|
| orion777 wrote: | | Maybe someone know, how to send shutdown command to the rpi3 if there are no ssh and attached usb keyboard? Because I have one raspberry as a NAT server, like "plug and it will boot", but how to stop them? |
Depending on what you're trying to do, you can Google for using the GPIO pins and a push button switch to provide a shutdown button. And indeed a start button that will boot from the shutdown state. Or 3-position toggle to do both. It's traditional on IBM mainframes for the power-off button to be huge and red and never used.
_________________
Greybeard |
|
| Back to top |
|
|
Goverp
Advocate
Joined: 07 Mar 2007
Posts: 2182
|
| Posted: Mon Aug 13, 2018 8:07 am Post subject: [EDIT] F2FS problem - or maybe not! |
|
|
[Edit]
Having reformatted the SD card to use ext4, it's showing the same errors, so it looks like the card is dead, rather then the filesystem. Ho hum; that'll teach me not to buy the cheapest cards.
[/Edit]
I've no-one to blame but myself, but, as f2fs worked OK for me on 32-bit Arch Pi3, when it came to 64bit Gentoo, I modified the installation to use it for the root partition. Generally seems to work, but an "emerge --update", something nasty happens after about 20 packages. Kernel messages and a semi-hang (can move among TTYs, but can't reboot or run any commands), specifically. I've updated to the latest kernel tarball; still happens. Just thought I'd warn people.
Anyway, many thanks for making all this available Sakaki.
_________________
Greybeard |
|
| Back to top |
|
|
Goverp
Advocate
Joined: 07 Mar 2007
Posts: 2182
|
| Posted: Wed Aug 15, 2018 11:29 am Post subject: |
|
|
Sakaki, a suggestion:
Would it be easy to build a squashfs image of /usr/portage as part of your weekly process? Emerge --sync updates a vast number of small files, exactly the wrong thing for an SD card. Downloading a new squashfs image would probably be quicker, and surely better for the card. And it would save lots of space, as the tree compresses really well, and it might reduce load on your rsync server.
To make it work sensibly, it would need moving the packages and distfiles directories out to somewhere in /var (possibly /var/cache), but they should have been there anyway if the early portage designers had thought for a bit.
_________________
Greybeard |
|
| Back to top |
|
|
Sakaki
Guru
Joined: 21 May 2014
Posts: 409
|
| Posted: Tue Aug 21, 2018 5:45 pm Post subject: |
|
|
| Goverp wrote: | Sakaki, a suggestion:
Would it be easy to build a squashfs image of /usr/portage as part of your weekly process? Emerge --sync updates a vast number of small files, exactly the wrong thing for an SD card. Downloading a new squashfs image would probably be quicker, and surely better for the card. And it would save lots of space, as the tree compresses really well, and it might reduce load on your rsync server.
To make it work sensibly, it would need moving the packages and distfiles directories out to somewhere in /var (possibly /var/cache), but they should have been there anyway if the early portage designers had thought for a bit. |
Thanks for the suggestion, I'll have a look into it. The problems are more likely to be making it reliable on the client (rather than the server) end I guess, as you'd need something to download and signature verify the image on a weekly sync, then dismount the old image, mount the new one (there'd be no point just using unsquashfs or similar, since that'd still have the small file writes), bind mount packages and distfiles into it and then clean up (optionally) past images that were too old. Do you know by any chance if anyone has written something similar for Portage in the past?
_________________
Regards,
sakaki |
|
| Back to top |
|
|
NeddySeagoon
Administrator
Joined: 05 Jul 2003
Posts: 54588
Location: 56N 3W
|
| Posted: Tue Aug 21, 2018 7:10 pm Post subject: |
|
|
Sakaki,
Would it help to start here for your weekly updates?
Move ./packages and ./distfiles out of /usr/portage, so you don't have the bind mount.
Then its throw one file away and install another.
Unless you did some auto updates to make.conf you would need a symlink pointing to the snapshot of the week, like linux points to a kernel whose name keeps changing.
eselect snapshot perhaps. That would allow for keeping several snapshots around at no extra charge.
_________________
Regards,
NeddySeagoon
Computer users fall into two groups:-
those that do backups
those that have never had a hard drive fail. |
|
| Back to top |
|
|
antonlacon
Apprentice
Joined: 27 Jun 2004
Posts: 257
|
| Posted: Wed Aug 22, 2018 1:02 am Post subject: |
|
|
There's a post sync hook you can use to build the squashfs tree by putting a script in portage/repo.postsync.d/ I have a sample here I run on my shared server back when I was serving up the portage tree to some clients here: https://github.com/antonlacon/gentoo/blob/master/portage/repo.postsync.d/squashfs-portage.sh Ignore most of the timestamp stuff (the timestamp comparison was to prevent building the same package when no sync actually took place, but had a few gaps in whether its test had the correct result), but the general approach might help you along.
There was also an old thread out of f.g.o's tips and tricks section about a squashfs portage tree and a unionfs overlay to enable syncing on top of it. I believe this was the initial thread: https://forums.gentoo.org/viewtopic-t-465367.html I had my own take on that approach here: https://github.com/antonlacon/gentoo/blob/master/initscripts/ (the squash-portage files in each directory). I'll offer that I haven't used said scripts in years (maybe a decade), and that looking at them now makes me think it's some fairly complicated and ugly shell, but maybe it's helpful anyway. I don't recall if a union fs ever got merged into the kernel. |
|
| Back to top |
|
|
|
|
You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot vote in polls in this forum
|
|
Gentoo on ARM
