Possibility to boot via PXE and load filesystem into ram?

[Chrysor]

Hello guys,

i have set up my Raspberry Pi 3 with PXE according to this tutorial: https://www.raspberrypi.org/documentati ... utorial.md

However, the problem with this is that it is only suitable for only one client, since there would be much space needed on the nfs server for different clients and because it would be causing heavy network load. That's why I would like to load only one file to the Pi containing the whole root filesystem which would then be inside some kind of ramdisk or something comparable, causing no more network load.

Is there any possibility to accomplish this? Has anyone done something like this before?
An SD card would absolutely not be an option for my purpose and due to the amount of clients an nfs server wouldn't be either.

best regards

[rpdom]

I don't see why it couldn't be done.

First build a system with an initrd, perhaps containing busybox. Put those files on the PXE server and you're done.

I've done similar with PCs in the past. I would try it out on a Pi, but I have too much to do already.

[Chrysor]

Hi,

thanks four answer. As far as I understood, initrd/initramfs just delivers the most basic filesystem, which the system needs to boot. How can I deliver the whole/the remaining filesystem to the Pi, making him store everything inside the ram, that would normally be on the sd card? I couldn't find any solution searching the web, but it is definitely done with desktop pcs like this. Maybe I just don't know what to search for

[rpdom]

You won't be able to store the whole filesystem in RAM. The full Jessie image won't fit on a 4GB card and you only have 1GB of RAM (minus whatever is assigned to the GPU).

A reduced system is the best you are going to get.

As far as I understood, initrd/initramfs just delivers the most basic filesystem, which the system needs to boot.

The standard ones do, but there is nothing to stop you making a more complete system that will provide some functionality.

Look at busybox for an example.

[Chrysor]

You won't be able to store the whole filesystem in RAM. The full Jessie image won't fit on a 4GB card and you only have 1GB of RAM (minus whatever is assigned to the GPU).

Yeah, I thought of that and thought of some small linux distributions. Would they still be too big because of the memory usage of the GPU?

[kusti8]

You won't be able to store the whole filesystem in RAM. The full Jessie image won't fit on a 4GB card and you only have 1GB of RAM (minus whatever is assigned to the GPU).

Yeah, I thought of that and thought of some small linux distributions. Would they still be too big because of the memory usage of the GPU?

The minimum for the GPU is 16 so if you have only command line, then it will be fine.

[Chrysor]

The minimum for the GPU is 16 so if you have only command line, then it will be fine.

Unfortunately I need to run a Web application inside a browser in some gui, so i think I'd maybe need around 128mb for gpu

[kusti8]

The minimum for the GPU is 16 so if you have only command line, then it will be fine.

Unfortunately I need to run a Web application inside a browser in some gui, so i think I'd maybe need around 128mb for gpu

Try looking at tiny Core Linux.

[asandford]

Iv'e found that PXE & iSCSI has less overhead and demand on network resources than NFS (I've been able to swap the cable from the onboard NIC to a USB gigabit adapter, and it just carried on).

[hazeii]

You can share most of the filesystem between all the clients, and keep just the variable bits in either local RAM or on a small per-client area on the server. With a gigabit network, it's not going to be troubled by 10 pi's at 100Mb each.

To keep it all 100% in local RAM, a crude but easy to use approach is to take an SD card, delete /usr and everything obviously not needed from /bin, /sbin and /lib. Once it's cut down to a reasonable size - and still boots - turn what's left into an initial ramdisk, get it booting and then add whatever apps you actually need.

Other approaches could include starting with the network installer and hacking it to suit your purpose, or some variant of the old-school hack on passing 'init=/bin/sh' on the kernel command line.

By the way, the boot isn't by way of 'PXE'; it only gets mentioned because the Pi sends out a packet saying it's a PXE client, but it's not really - it always uses 'bootcode.bin' (even if the network server tells the Pi it should use a PXE program as a result, the Pi ignores it and requests 'bootcode.bin').

[rln]

and thought of some small linux distributions. Would they still be too big because of the memory usage of the GPU?

The development version of my Nard SDK might work. It's not entirely finished yet, but it can boot into a RAM standalone system. It's console based, but there is a framebuffer where one can display JPEG images etc.
http://www.arbetsmyra.dyndns.org/nard/

[Chrysor]

Iv'e found that PXE & iSCSI has less overhead and demand on network resources than NFS (I've been able to swap the cable from the onboard NIC to a USB gigabit adapter, and it just carried on).

You can share most of the filesystem between all the clients, and keep just the variable bits in either local RAM or on a small per-client area on the server. With a gigabit network, it's not going to be troubled by 10 pi's at 100Mb each.

Maybe I should've mentioned that I'm talking about theoretical amounts of devices above 1000. You might see that using any network based file system would clearly cause too much traffic. Furthermore, since Raspberries tend to destroy SD Cards from time to time, using them would cause another problem I wouldn't want to deal with.

To keep it all 100% in local RAM, a crude but easy to use approach is to take an SD card, delete /usr and everything obviously not needed from /bin, /sbin and /lib. Once it's cut down to a reasonable size - and still boots - turn what's left into an initial ramdisk, get it booting and then add whatever apps you actually need.

I'll give that a try, but I haven't been able to find out how to create such an initial ramdisk using for example mkinitramfs. Can you tell me which tool to use for this?

Thanks so far!

[Chrysor]

Hey guys,

so I managed to create an initial ramdisk image, but I'm unsure how to proceed now in order to test it. When I boot my Pi from network, it loads the bootloader and more files including the cmdline.txt file, where it is specified, that /dev/mmcblk0p2 is the root device. This is untrue of course and must be changed. But I have no clue to what it needs to be changed. Let's say I'd change it to /dev/ram0 or sth like that, then there still must be something mounted there already, right?
I also put the line "ramfsfile=initrd.img" inside the config.txt file, but it seems like the Pi wants to mount the root filesystem first, because I'm getting this output line while booting: "Waiting for root device /dev/mmcblk0p2..."

Any ideas?

[hazeii]

You actually don't need the 'root=' argument if you're going to run everything from the (initial) ramdisk. Basically the kernel will pick up the initial ramdisk from the entry in config.txt , load that as the initial root filesystem and (by default) try to run /sbin/init (which can be a shell script - see 'initrd.txt' in the kernel docs for details).

If you don't actually have anything set to run on the initial ramdisk, the kernel will continue as normal (with the ramdisk as the root filesystem). As long as you have a shell in your ramdisk, add "init=/bin/sh" to cmdline.txt and the kernel will run that.

A very simple ramdisk can contain nothing but a single program (together with any libraries it depends on). In the old days there would always be a statically linked shell in /bin, so 'init=/bin/sh' could be used to recover the system.

What it sounds like you have is the initial ramdisk loading but with nothing running from it, so the kernel continues (and tries to mount the SDCARD filesystem).

Here's my cmdline.txt (the 'ip=dhcp' isn't needed if you don't want networking):

Code: Select all

dwc_otg.lpm_enable=0 console=tty1 ip=dhcp logo.nologo rw

[rodrigoreimberg]

Hello,


Could you build the image and use the pxe boot? it's working?


Tks,

Rodrigo

[rln]

This is an old thread, but I can answer.

The answer is no. PXE support is way to buggy in the bootloader. One can use it on a lab bench. But for any serious mass deployment it's way to unreliable.