Marc's Public Blog - Linux Hacking

you can see the puffy capacitors and one of them died and took the motherboard with it

First, I tried to see if I could move all this to another VA Server I had laying around at home, a mighty dual P3 800Mhz. It would have been a downgrade, but would have brought the machine back up:

However, I quickly found out that the machine was way too old, and incapable of booting from Sata when a sata card was plugged in, or anything plugged via USB. Maybe the fact that the motherboard was from the late 90's had something to do with it :)
This time, I thought I'd be smarter and use the 2 unused Dell Poweredge 2950 servers I had sitting at home for a sunny day (maybe for up to 10 years. This time, I'd use a proper server and it would be easier :)

Also, I would finally upgrade to a 64bit kernel (my previous servers could only run 32bit, which started to be a problem with some software that was 64bit only, badly tested on 32bits, never mind code that needed more than 2GB of RAM (I mean technically you could run more than 2GB of RAM on a 32bit kernel, but it was through bounce buffers and non contiguous memory, with still a 2GB limit for each user space process).

But of course, I didn't look up that those servers dated from 2006 and were totally obsolete in 2019. I figured, it wouldn't matter, they were better than the previous even more obsolete free motherboard a coworker gave me and that I managed to fit in a VA Linux server case. Also, they had serial console and even remote serial console over network via a separate IP, luxury!!!

It also had nice disk hardware for SATA drives, so things were going to be great. I however found out that the required Perc Raid card (required because of the backplane the drives were plugged into and cabling), did allow some kind of Sata passthrough, so that I could do software raid with linux (I've never trusted hardware raid, it's always been too slow and vendor dependent). Sadly, it was terribly slow. I then went on the internet a bought some upgraded H700 Raid cards that were supposed to be faster. First I found out they needed new cables, sigh, but even once I got all of this, the entire system disk I/O was still unbearably slow.
In the end, with time running out (days had gone by while I had the old drives running off a server shell on my desk at home and serving queries and Email)

So this was the new server, looked nice inside:

Unfortunately I was forced to use the weird and slow raid card, which was still slow after upgrading to an H700:

4 Cores, 16GB of RAM, luxury in 2006 :)

Did I say it was slow? Here is a raid rebuild. I think reads were ok, but writes were terribly slow probably because they were forced to go through some writeback cache on the raid card that I did try to remove. Yeah, this first rebuild shows about 1 week for 7TB :-//

md3 : active raid5 sda6[7] sde6[6] sdf6[4] sdd6[8] sdb6[9] sdc6[2]
      7118325760 blocks super 1.2 level 5, 512k chunk, algorithm 2 [6/5] [_UUUUU]
      [=======>............]  recovery = 42.9% (611422712/1423665152) finish=3325.8min speed=4070K/sec

Interestingly, if I rebuild the array on the 7th drive, connected directly via Sata, it's 10x faster:

md1 : active raid6 sdg3[9] sda3[6] sde3[5] sdd3[7] sdb3[8] sdc3[2]
      419164160 blocks super 1.2 level 6, 512k chunk, algorithm 2 [6/5] [UUUU_U]
      [======>.............]  recovery = 38.7% (40655068/104791040) finish=19.8min speed=53705K/sec
      bitmap: 1/1 pages [4KB], 65536KB chunk

The previous server was really built on happy thoughts and good luck, it's amazing it ran for 11 years and I didn't have any major double disk failure (I did have some drives I had to replace over time, but nothing major and never an emergency or with any downtime). This time around, I did things seriously:

I really really wanted to just put a big sata card and plug it onto all the drives, but the Dell backplane and drive power routing made it mostly impossible. So after the raid card upgrade, I realized it was a lost cause, and changed plans to boot from a 4TB SSD and use it as the main boot and data drive:

you can see the 7th drive sitting on top of the motherboard and using a spare sata port, and the 4TB SSD on top

Now, the 4TB Sata SSD, which I got for free after one in a set died an early death, taking all its data with it, did not fill me with confidence, but it was fast and fixed my slow raid issues, so I made the call to use it as a boot drive. Thanks to my btrfs send/receive snapshots setup, the SSD was backed up hourly to spare filesytems on the raid, and it would be easy to switch over remotely with virtually no data loss if it died.
I found out in Oct 2025 that it survived 6 years without dying. I still wouldn't trust it much, but it survived, so good enough :)

As laughable it was to setup what turned out to be a 2006 server in 2019, and putting aside the issues with very slow write speed through the raid card I never solved, this was the most overbuilt system I ever built

What filesystems looked like:

sdh4          400G  0 part  /mnt/btrfs_boot
sdh5          1.6T  0 part  /mnt/btrfs_pool1
md0           953M  0 raid1 /boot2 (spare ext4 linux boot/recovery system)
md1         399.8G  0 raid6 /mnt/btrfs_bootb (btrfs send/receive of sdh4)
md2           1.6T  0 raid6 /mnt/btrfs_pool1b (btrfs send/receive of sdh5)

Quad boot system:

/dev/sdi1          63   24191   24129 11.8M  1 FAT12
/dev/sdi2       24192   48383   24192 11.8M  1 FAT12
/dev/sdi3  *    48384 4011839 3963456  1.9G  b W95 FAT32

/dev/sdh1        2048       4095       2048    1M BIOS boot
/dev/sdh2        4096    2111487    2107392    1G EFI System
/dev/sdh3     2111488   39063551   36952064 17.6G Linux extended boot
/dev/sdh4    39063552  877924351  838860800  400G Linux filesystem
/dev/sdh5   877924352 4283699199 3405774848  1.6T Linux filesystem
/dev/sdh6  4283699200 4317253631   33554432   16G Linux swap
/dev/sdh7  4317253632 4585689087  268435456  128G Intel Fast Flash
/dev/sdh8  4585689088 7501476494 2915787407  1.4T Linux filesystem

Then 2 more raids with data:

This is actually the only version of magic I built and that didn't die. I only replaced it with, ironically enough, a duct taped frankenstein Raspberry Pi 5 that really doesn't look solid, but that I hope will work as I write this in 2025. The reason for the switch, is it uses 20x less power and is almost 3x faster, so it's pretty much 60 times more efficient and takes the power use from 400W-ish to 15W

Price of Power use

Price of power is no joke:

Silicon Valley Power (Santa Clara): $0.175/Kwh	$613/year
Palo Alto Utilities (Palo Alto): $0.22/Kwh	$771/year
Pacific Gas & Electric (PG&E): $0.425/Kwh	$1,489/year

Further reading