Jim's Depository

this code is not yet written

I needed a machine to do some DNS server tests. I settled on a \$280 EEE PC 900A (stripped of webcam and half of its storage) from Best Buy. That gets me a 1.6GHz x86 server with 1G of ram that only burns 10 watts and comes with its own little console for when I need it. Not a bad deal. 

Only 4G of storage, but I’m only using about 60% even with a bunch of heavy eyecandy gnome and compiz stuff I installed to see what would happen (it is pretty fast, lower end graphic accelerator, but not many pixels comes out well).

I wiped the friendly linux it came with and installed Debian Lenny and all is good, except I kept noticing intermittent disk hangs lasting several seconds. I think I finally tracked this down to the kernel syncing out written pages. The fix is to not write so much. By mounting the partitions noatime most of my writes go away and I don’t notice hangs anymore.

Reading the first byte of every file in /usr went from 131 seconds to 92 seconds with the change (after a fresh boot each time), that is about a 30% speedup.

I’m pleased with the EEE. My code builds from clean in 1.6 seconds. I rarely use more than 10% of the RAM doing development which leaves plenty of RAM for caches to mitigate the slow flash disk. 

At last, I can put my /boot partition in LVM.

  • Get the Debian box up to Lenny.
  • Note that I accidentally trashed my MBR and had to boot into rescue mode while working out these steps. You shouldn’t do this if you follow all of the instructions, but you ought to have media handy.
  • aptitude install grub-pc (Note: this will remove the old grub package and offer to chain load grub2 from your existing grub. Do this. If you have problems you can still boot.)
  • Verify you can reboot.
  • Remove the old grub MBR and put in the grub2 one with upgrade-from-grub-legacy
  • Hide your /boot/grub/menu.lst so you aren’t tempted to edit it.
  • Your basic configuration, like kernel command line parameters is now in /etc/defaults/grub, there is also /etc/grub.d/* which I hope to never touch.
  • Move your /boot into the LVM. You could tar up your /boot partition, unmount it, and extract it onto the root partition. You could also make a new LVM managed /boot if you like it on its own partition. I was out of space in the volume groups, so I went with /. If you didn’t make a new /boot, remember to take /boot out of /etc/fstab.
  • dd zeros onto your old boot partition to make sure you aren’t deluding yourself.
  • Edit /etc/defaults/grub to add GRUB_PRELOAD_MODULES=lvm
  • Go back and make certain you did the previous step. I made an unbootable system before I learned that little tidbit.
  • Do an update-grub and a grub-install /dev/sda or whatever your disk is.
  • Go back and make sure you did the grub-install… Just update-grub is not enough to pick up the lvm module.
  • Reboot and rejoice.

I am left wondering what silliness lead to GRUB-2 being version 1.96, but I am happy.

You saved me many (more) hours of head pounding with this blog entry.  I am thoroughly grateful.

At some point in the past I managed to screw up my file server's lenny install in such a way that I ended up with the non-lvm ext2 boot partition commented out of fstab and a separate /boot directory on the lvm root.

I forgot about this incident and went about continuing to run apt-get dist-upgrade periodically.  Everything worked until I went to squeeze and rebooted, at which point I made some more poor choices ("Why am I not running the new kernel?  I'll just apt-get remove the old one!") and ended up unable to mount ext2 partitions (while still able to boot from one).

After about eight hours of head scratching I found this page and by following your steps had no trouble upgrading to GRUB 2 which booted the new kernel which fixed all the problems, allowing me to get on with my life (such as it is).

You are awesome and so is GRUB 2.

You should note that grub-pc for lenny is missing part_msdos.mod and will not install, you need to fetch it from backports to get the file.

Thanks for the info though, it helped me to confirm what I was doing would work before I sent a remote machine through a reboot (I know, dangeous, but couldn't be helped)
I have been trying to do this for 3 whole days and I finally found this post by accident. I have tried to do this with Arch Linux and Fedora and had no success. Even Google yielded no answers until I tried "Debian grub2 lvm /boot" and found this. Even though this is not written for squeeze I managed to take the bottom half and make this work. I even went as far as asking on 2 forums and in IRC channels for multiple OS's, even bugged some friends and got the usual "why would you want to do that?". Sir, I cannot thank you enough. Looks like I'm a Debian user now thanks to blogs like this.
Debian Squeeze, VirtualBox 4.0 from backports.

Debian 6.0.4 installer .iso

6 virtual disk devices
-> one physical raid partition per device
-> one RAID6 md0

One physical volume group
logical volume /
logical volume /home

The system does not manage to boot even after going in into the rescue mode and making sure grub is both configured and installed with the lvm module preloading (/etc/default/grub -> update-grub -> grub-install).

It all trips up in grub-pc (GRUB2) somehow and I do not know how to debug it.

GRUB loading.
Welcome to GRUB!

error: file not found.
Entering rescue mode...
grub rescue>

Any pointers? Did I miss and mess up any of the rescue mode superhero stuff?
I tried to wave a rubber chicken and just throw every plausibly connected module at it, but it's still a no-go.

I've seen some indirect talk about mdraid + LVM + EXT4 not being a bootable combo yet, but then again I've seen people boast about their mdraid + dmcrypt + LVM + EXT4 / BTRFS setups.

I guess I will just fall back to a good old fashioned EXT2 boot partition.

Here's my rubber chicken waving approach for the internet:

GRUB_PRELOAD_MODULES="search_fs_uuid raid raid5rec raid6rec mdraid lvm ext2 chain pci"


grub-install --modules="search_fs_uuid raid raid5rec raid6rec mdraid lvm ext2 chain pci" /dev/sda

Little known and not terribly documented, but true:

You can tell your Debian machine to do an apt update and and download any files that you might need by adding two lines to your apt config (which I’ll bet you didn’t know you had.)

cat > /etc/apt/apt.conf.d/50autoupdate APT::Periodic::Update-Package-Lists “1”; APT::Periodic::Download-Upgradeable-Packages “1”;

Then the script /etc/cron.daily/apt will keep you all up to date savings you seconds to minutes every time you decide to upgrade.

Another good one: To use a cacheing proxy for your packages, add this line to a similar file:

Acquire::http::Proxy "http://YOURHOST:YOURPORT";

This way you don't have to mess with your /etc/apt/sources.list file to make all the proxy changes.

Cacheing common reports from a database inside the database itself is surprisingly easy and makes a huge difference for Ajax style web pages.

I have several systems that record data at regular intervals. For the sake of example let us consider a weather station which reports temperature, humidity, pressure, and wind every 10 minutes. If I want to graph this data for a three day period, I have to query out 432 of these rows and send them down to my browser. Unfortunately this is not fast enough.

Step 1: Get the browser cache working for me.

If I break the request in to midnight aligned 24 hour periods, then I can cache the result for any completed day. This way I only need to pull new days of data. 

This helps, but it turns out I don’t revisit days often during a session.

Step 2: Tune the database indices.

Fail miserably. It turns out SQLite does my query faster without indicies, so I took them off completely. (Sequential read on my virtual server is much faster than random.)

Step 3: Server side cacheing. In the database.

Now we get to the meat. I can cache pre-compressed reports for each of my daily periods. There are a couple of wrinkles though. I need some way to invalidate a report when the underlying data changes. (Sometimes some observations can be delayed and trickle in later.) I can’t think of a good way to have the database delete a cache file, so instead I store the cached copies in the database.

This turns out to be surprisingly clean to code, and a simple set of triggers on the underlying data can remove any affected report.

It would be more efficient to keep the cached reports as files, or better let in a httpd cache, but then they could not be invalidated by the database. 

Step 4: Clean up the little PHP wrapper and stick it here.

I’ll get to that.