Ceph Cluster Updates

The Ceph cluster is continuing to run fine and generally meets all of my needs, though I could definitely use more IOPs, especially for RBD volumes attached to VMs. I have encountered a couple operational issues that may crop up in a home environment outside a datacenter with redundant internet connectivity and power. If you’d like to read more on the cluster hardware choice, setup and testing, you can here.


I had a nearly 3-day Internet outage caused by a Telus’ general sloppiness. While I was offline, my cluster wasn’t able to keep time using NTP. While I had the cluster computers pulling NTP from a local NTP server (running on my PFSense router), all of its sources of time were on the Internet, except that derived from its onboard clock. At some point, NTP on the Ceph servers decided that the upstream was bad, and so dropped it. Over a day the clocks on the servers drifted enough that Ceph started having serious issues. I ended up needing to shut down the cluster, and with it, all the VMs using it for storage, because it had degraded to the point of being unusable due to how Ceph handles clock skew.
Luckily, I had a USB based GPS unit handy, and I was able to set myself up a backup stratum 1 NTP server using the GPS and a Raspberry Pi. While USB GPS units generally have significantly more jitter than serial ones due to the vagaries of USB’s timing, it was much lower than the clocks on the servers, and this satisfied Ceph.With higher quality timekeeping my cluster was back to being healthy and able to serve writes.
I ended up using GPSD to access the GPS and Chrony to handle being the NTP server. I found this guide and this one useful to get things working.

Lessons learned: Time synchronization is important for a healthy Ceph Cluster. Have a local backup time source. While Ceph can be tuned to be more tolerant to poor time sync, there’s no replacement for stable time infrastructure.


A month or so after the Internet outage, I had a power outage that lasted 1.5 hours. My backup power for the cluster has about 75 minutes of runtime, so the cluster was shut down. I powered down the VMs and then set the following flags on the cluster (source):

ceph osd set noout
ceph osd set nobackfill
ceph osd set norecover

To bring the cluster back, I started the three hosts back up and unset the flags. However, when everything came back up, I realized that I hadn’t waited enough time before powering down the hosts, and some of the VM’s RBD disks hadn’t finished syncing. Cue several hours of running xfs_repair on a bunch of RBD devices. In the end, I didn’t lose any data, just had lightly corrupt filesystems.

Lessons learned: Safe shutdown needs to be handled automatically, with attention taken to how the cluster is being used. I wrote a safe shutdown script to use with Proxmox and apcupsd that shuts down the VMs, calls sync twice and then waits a (hopefully very conservative) 10 minutes before powering off the hosts, to ensure that all data has been properly written to disks. As my hosts are not set to power back on when power comes back, I’m manually un-setting the flags when the system come up.

Other Tweaks

Sometime I was getting Ceph into HEALTH_WARN with showing OSDs with slow ops, and dmesg was full of errors like nvme nvme0: I/O 42 QID 47 timeout, completion polled. This appears to be due to a firmware bug in the Intel P4510 SSDs I’m using, that has since been fixed. So I did some baseline tests and then updated the firmware on the hosts and rebooted.

  • Download the Intel Memory and Storage Tool (MAS), CLI version for Linux from Intel’s website. Strangely, Proxmox doesn’t have this packaged (licensing issues).
  • Unzip it.
  • Install the debian package using apt install ./intelmas_<version>_amd64.deb.
  • Find the SSD using MAS: intelmas show -intelssd
  • Update the firmware: intelmas load -intelssd <SSD #>.
  • Reboot

This update tripled the I/O throughput of the nvme-only pool residing on the SSDs, both in IOPs and raw bitrate. Maximum latency, as reported by sysbench went from 30 seconds to 90 milli-seconds. And, no more errors in dmesg or slow ops on my OSDs.


I’m still pretty happy with Ceph at home, but there are certainly some considerations that crop up with home use that don’t in a datacenter, though they should probably also be planned for even with redundant power and connectivity, in the event of a true disaster.

Below is my current cluster usage. I’ve got a 37TiB CephFS filesystem running on it, with the remainder being used for RBD images for VMs and the like.

    id:     5afc28eb-61b3-45f1-b441-6f913fd70505
    health: HEALTH_OK

    mon: 3 daemons, quorum megaera,tisiphone,alecto (age 5w)
    mgr: alecto(active, since 5w), standbys: megaera, tisiphone
    mds: cephfs-ec:1 {0=tisiphone=up:active} 2 up:standby
    osd: 21 osds: 21 up (since 5w), 21 in (since 3M)

    pools:   4 pools, 320 pgs
    objects: 23.78M objects, 35 TiB
    usage:   57 TiB used, 44 TiB / 101 TiB avail
    pgs:     320 active+clean

Three Node Ceph Cluster at Home

I’ve always wanted to use Ceph at home, but triple replication meant that it was out of my budget. When Ceph added Erasure Coding, it meant I could build a more cost-effective Ceph cluster. I had a working file-server, so I didn’t need to build a full-scale cluster, but I did some tests on Raspberry Pi 3B+s to see if they’d allow for a usable cluster with one OSD per Pi. When that didn’t even work, I shelved the idea as I couldn’t justify the expense of building a cluster.

When my file-server started getting full, I decided to build a Ceph cluster to replace it. I’d get more redundancy, easier expansion and have refreshed hardware (some of my drives are going to be 9 years old this summer). I briefly looked at ZFS. But between its limited features and with the legality of running it on Linux being an open question, I quickly ruled it out.

The Cluster – Hardware

Three nodes is the generally considered the minimum number for Ceph. I briefly tested a single-node setup, but it wasn’t really better than my file-server. So my minimal design is three nodes, with the ability to add more nodes and OSDs if and when my storage needs grow.

A stack of boxes with computer hardware in them, waiting to be built into a storage cluster, sitting on a dining room table.
Some of the hardware waiting to be built into cluster nodes sitting on my dining room table. Not pictured is the first node I built for a single-node test.

I wanted each node to be small and low-ish power, so I was mainly looking at mATX cases that could take 8 3.5″ drives. After much searching, I realized that mATX is basically dead and there aren’t many mATX cases or motherboards out there. Presumably people either get mITX cases and motherboards if they want something small or get full ATX boards if they want lots of on-board peripherals and expandability, leaving mATX as an awkward middle ground.

For the motherboard, it needed to have (either onboard or space to add via add-in cards): 8xSATA (or SAS), 2x M.2 (at least one 22110) slots, 8-core CPU (or better), support for 32GB RAM sticks and a single 10GbE port (not SFP+). I found three boards that would work, a SuperMicro board that was incredibly expensive but had everything except the CPU onboard, an AMD X570 based board and an Intel Z390 based board. I quickly ruled out the SuperMicro board based on price and the Intel board based on the lack of low-power CPUs (to get the 8-cores, I was looking at a 9700k, which isn’t low-power or particularly fast or the 9900T, which no one could get me). I chose an ASRock X570M Pro4 board. It had what I needed, and better yet, it supported a 65W, 8-core Ryzen 3 CPU. I’d been bit by serious hardware bugs in the Ryzen 1, so I was a bit wary to try it, but Intel had nothing anywhere near competitive.

HDD choice was relatively easy, I created a spreadsheet with all the drive models I could find, found the best $/TB 7200RPM model I could. This ended up being 6TB Seagate Ironwolf drives, which were discontinued in favour of an inferior model after I ordered mine. Luckily with Ceph, replacements don’t need to be the same size like they did in RAID6.

SSDs were a little more challenging. I chose inexpensive, but good, M.2 SSDs for the OS drives. I also wanted some SSD based OSDs. Ceph apparently does not do well on consumer level SSDs without power-loss-protection and consequently has very slow fsync performance, so I needed a fancier SSD than the Samsung 970 Pros I had been intending to use. I found the Intel P4510/P4511 series, and decided on a 2.5″ U.2 P4510. This required an M.2 to min-SAS adapter and mini-SAS to U.2 adapter to get it connected to the board’s open M.2 slot. Why not use the P4511? No stock on it.

Fractal Design Node 804 – Case38×3.5″, 2×2.5″, mATX, full ATX PSU
ASRock X570M Pro4 – Motherboard3mATX, 8xSATA, 2x M.2, PCIe4.0 x16, x1, x4
Corsair RM550x – Power Supply3ATX PSU
Seagate 6TB Ironwolf – Hard Drive18OSD: 6 per node, 108TB Raw, ST6000VN0033
Kingston SC2000 250GB – M.2 SSD3OS/Boot Drive
Intel P4510 1TB – U.2 SSD3OSD: 1 per node
AMD Ryzen 3700x = CPU365W, 8-core
64GB Corsair LPX – RAM3One 2x32GB DDR4 3200 per node
Geforce GT710 – Video Card3One per node
Startech M.2 to U.2 Adapter Board3To connect P4510 SSD to motherboard
Mini-SAS to U.2 Adapter Cable3Cable to connect from Startech adapter to SSD
4x SATA Splitter Cable6One per each bank of 4 drives, 2 per node
Corsair ML120 120mm 4-pin fan3One each on front of drive compartment
Aquanta AQ-107 – 10GbE NIC3One per node
Cat 6a Patch Cable3One per node
Netgear 8-port 10GbE Switch1Model: XS708E, one for the whole cluster
Cluster hardware.

A small note on networking: I elected not to have separate public and cluster networks, I set everything to use the same 10GbE network. This simplified setup, both on the host/Ceph as well as physical cabling and switch setup. For a small cluster, the difference shouldn’t matter.

An Ikea Omar rack with three computers as well a UPS at the bottom, two network switches at the top and a bunch of network cabling.
Three cluster nodes in an Ikea Omar wire rack. At the bottom is a 1500VA APC UPS with a 3kVA additional battery. At the top is my core switch, and the cluster’s 10GbE switch.

Other hardware notes: the Fractal Design Node 804 HDD mounts are missing one of the two standard screw holes. The spec for 3.5″ drives only requires the two end holes, but the Node 804’s mounts only support the optional middle hole and the hole nearest the connectors. Most drives 6TB and over lack the middle hole, apparently to allow another platter.

The Cluster – Setup

Setup was pretty straightforward. I used Arch Linux as a base, running Ceph version 14.2.8 (then current). I installed the cluster using the Manual Deployment instructions.

Once the cluster was running, it was time to create pools and set up the CephFS filesystem I planned on migrating to. Ceph correctly assigned my hard drives as hdd class and the SSDs as ssd class. I had planned to have CephFS backed by an erasure coded pool, with a durability requirement of being able to lose either two drives or one host (but not both). CephFS doesn’t allow EC coded pools to be used to the CephFS metadata store, so I created a replicated pool on the P4510 SSDs. I’m using the P4510s to store the metadata because suggestions online were that it would increase performance to have the metadata pool on SSDs. I’m not sure if this actually makes much difference with the low number of drives I have.

I created a CRUSH rule that will place data only on SSDs with: ceph osd crush rule create-replicated ssd-only default host ssd.

To make sure this CRUSH rule worked, I tested it by:

  1. Dumping a copy of the crushmap using ceph osd getcrushmap -o crushmap.orig
  2. Running crushtool --test -i crushmap.orig --rule 2 --show-mappings --x 1 --num-rep 3 (the number after --rule is the index of the rule to test).
    Results: CRUSH rule 2 x 1 [18,20,19], which are the OSD numbers of my SSDs, exactly as intended.

Finally, create the pool with ceph osd pool set cephfs-metadata crush-rule-name ssd-only. Excellent! On to the EC pool.

Three Node Cluster – EC CRUSH Rules

The EC coded pool took a little more work to get working. My design goal is to have the cluster be able to suffer the failure of either a single node or two OSDs in any nodes. To do this, I would minimally need to split each block up into four pieces, plus two parity pieces.

What the OSD tree looks like:

$ ceph osd tree
-1       100.97397 root default
-3        33.65799     host Alecto
 0   hdd   5.45799         osd.0          up  1.00000 1.00000
 1   hdd   5.45799         osd.1          up  1.00000 1.00000
 2   hdd   5.45799         osd.2          up  1.00000 1.00000
 3   hdd   5.45799         osd.3          up  1.00000 1.00000
 4   hdd   5.45799         osd.4          up  1.00000 1.00000
 5   hdd   5.45799         osd.5          up  1.00000 1.00000
20   ssd   0.90999         osd.20         up  1.00000 1.00000
-5        33.65799     host Megaera
 6   hdd   5.45799         osd.6          up  1.00000 1.00000
 7   hdd   5.45799         osd.7          up  1.00000 1.00000
 9   hdd   5.45799         osd.9          up  1.00000 1.00000
11   hdd   5.45799         osd.11         up  1.00000 1.00000
14   hdd   5.45799         osd.14         up  1.00000 1.00000
16   hdd   5.45799         osd.16         up  1.00000 1.00000
19   ssd   0.90999         osd.19         up  1.00000 1.00000
-7        33.65799     host Tisiphone
 8   hdd   5.45799         osd.8          up  1.00000 1.00000
10   hdd   5.45799         osd.10         up  1.00000 1.00000
12   hdd   5.45799         osd.12         up  1.00000 1.00000
13   hdd   5.45799         osd.13         up  1.00000 1.00000
15   hdd   5.45799         osd.15         up  1.00000 1.00000
17   hdd   5.45799         osd.17         up  1.00000 1.00000
18   ssd   0.90999         osd.18         up  1.00000 1.00000

I found a blog post from 2017 describing how to configure a CRUSH rule to make this happen, but that post was a little light on how to actually do this.

  1. Get a copy of the existing crushmap: ceph osd getcrushmap -o crushmap.orig
  2. Decompile the crushmap to plaintext to edit: crushtool -d crushmap.orig -o crushmap.decomp
  3. Edit the crushmap using a text editor. This is the rule I added:
rule ec-rule {
        id 1
        type erasure
        min_size 6
        max_size 6
        step set_chooseleaf_tries 5
        step set_choose_tries 100
        step take default class hdd
        step choose indep 3 type host
        step choose indep 2 type osd
        step emit

Some explanation on this rule config:

  • min_size and max_size being 6 is how many OSDs we want to split the data over.
  • step take default class hdd means that CRUSH won’t place any blocks on the SSDs.
  • The lines step choose indep 3 type host and step choose indep 2 type osd tell CRUSH to first choose three hosts and then CRUSH to choose two OSDs on each of those hosts.

4. Compile the modified crushmap: crushtool -c crushmap.decomp -o crushmap.new

5. Test the new crushmap: crushtool --test -i crushmap.new --rule 1 --show-mappings --x 1 --num-rep 6.
In my case, this resulted in CRUSH rule 1 x 1 [9,16,8,13,5,0], which shows placement on 6 OSDs, with two per host.

6. Insert the new crushmap into the cluster: ceph osd setcrushmap -i crushmap.new

More information on this can be found on the CRUSH Maps documentation.

With the rule created, next came creating a pool with the rule:

  • Create an erasure code profile for the EC pool: ceph osd erasure-code-profile set ec-profile_m2-k4 m=2 k=4. This is a profile with k=4 and m=2, so two parity OSDs and 4 data OSDs for a total of 6 OSDs.
  • Create the pool with the CRUSH rule and EC profile: ceph osd pool create cephfs-ec-data 128 128 erasure ec-profile_m2-k4 ec-rule. I chose 128 PGs because it seemd like a reasonable number.
  • As CephFS requires a non-default configuration option to use EC pools as data storage, run: ceph osd pool set cephfs-ec-data allow_ec_overwrites true.
  • The final step was to create the CephFS filesystem itself: ceph fs new cephfs-ec cephfs-metadata cephfs-ec-data --force, with the force being required to use an EC pool for data.


Once that was all done, I finally had a healthy cluster with two pools and a CephFS filesystem:

    id:     5afc28eb-61b3-45f1-b441-6f913fd70505
    health: HEALTH_OK

    mon: 3 daemons, quorum alecto,megaera,tisiphone (age 11h)
    mgr: alecto(active, since 11h), standbys: megaera, tisiphone
    mds: cephfs-ec:1 {0=alecto=up:active} 2 up:standby
    osd: 21 osds: 21 up (since 11h), 21 in (since 11h)

    pools:   2 pools, 160 pgs
    objects: 22 objects, 3.9 KiB
    usage:   22 GiB used, 101 TiB / 101 TiB avail
    pgs:     160 active+clean

Instead of running synthetic benchmarks, I decided to copy some of my data from the old server into the new cluster. Performance isn’t quite what I was hoping for, I’ll need to dig into why, but I haven’t done any performance tuning. Basic tests were getting about 40-60MB/s write with a mix of file sizes from a few MB to a few dozen GB. I was hoping to max my fileserver’s 1GbE link, but 60MB/s of random writes on spinning rust isn’t bad, especially with only 18 total drives.

If you’re curious about the names I chose for my 3 hosts, Alecto, Megaera and Tisiphone are the names of the three Greek Furies. If I add more hosts, I’m going to be a bit stuck for names, but adding the Greek Fates should get me another 3 nodes.

One final note: when I was trying to mount CephFS I kept getting the error mount error: no mds server is up or the cluster is laggy which wasn’t terribly helpful. dmesg seemed to suggest I was having authentication issues, but the key was right. Turns out that I also needed to specify the username for it to work.

Raspberry Pi Ceph Cluster – Testing Part 2

Having built a Ceph cluster with three Raspberry Pi 3 B+s and unsuccessfully tested it with RBD, it’s time to try CephFS and Rados Gateway. The cluster setup post is here and the test setup and RBD results post is here. Given how poorly RBD performed, I’m expecting similar poor performance from CephFS. Using the object storage gateway might work better, but I don’t have high hopes for the cluster staying stable under even small loads in either test.

Test Setup:

I’m using the same test setup as I used in the RBD tests. Two pools, one using 2x replication and the other using erasure coding. Test client is an Arch Linux system running a 5.1.14 kernel with Ceph 13.2.1, a quad core CPU, 16GiB of RAM and connected via a 1GbE connection to the cluster. I’m also running the OSDs with their cache limited to 256MiB maximum size, and the metadata cache limited to 128MiB.


CephFS requires a metadata pool, so I created a replicated pool for metadata. Why not create both at once? CephFS doesn’t currently support multiple filesystems on the same cluster, though there is experimental support for it.

With the pool created and the CephFS filesystem mounted on my test client, I started the dd write test with a 12GiB file using dd if=/dev/zero of=/mnt/test/test.dd bs=4M count=3072. The first run completed almost instantly, apparently completely fitting in the filesystem cache. CephFS doesn’t support iflag=direct in dd, so I simply reran the write test knowing that the cache was pretty full at this point. Almost instantly, with less than 1GiB into the test, two of the nodes simultaneously fell over and died. They were completely unreachable over the network, but this time I connected a HDMI monitor to them to see the console. I saw quickly that kthread had been blocked for over 120 seconds, and the system was pretty completely unusable. A USB keyboard was recognized, but key-presses weren’t registering. I power-cycled the systems after waiting at least five minutes, and they came up fine.

A very unhappy Ceph cluster with two hosts that have locked up.

I tried running the test again, and both hosts quickly locked up. I was running top as they did so, and both hosts rapidly consumed their memory. Despite having the OSD’s cache set to a maximum of 256MiB, the ceph-osd process was using around 750MiB before the system became unresponsive. OOM killer didn’t kill cpeh-osd in this case to save the system, possibly due to the kernel failing to allocate memory internally to do so. CephFS seems to hang the Raspberry Pis hard.

I decided to test a bunch of smaller files, because CephFS is meant more as a general filesystem with a bunch of files, whereas RBD tends to get used to store large VM images. I used sysbench to create 1024 16MiB files for a total of 16GiB on my 40GiB CephFS filesystem. Initially, things seemed to work fine. Sysbench reported that it created the 1024 files at 6.4MB/s. While this was just test preparation, it seemed to be a good sign.

What didn’t seem such a good sign was when I actually started running the sysbench write test and Ceph started complaining about slow MDS ops. A lot of them. The sysbench write test immediately failed, citing an IO error on the file. Running ls -la showed a lot of 0 bytes file, with a couple 16MiB files. Ugh. I recreated the test setup, this time with writes at a blazing 540kB/s. When it finally finished several hours later, attempting to run the write tests showed the same truncation of files to 0B as before. This seemed to be a sysbench issue, but I didn’t spend much time troubleshooting it.

For completeness, I also tried an erasure coded pool with CephFS. Like RBD, the metadata pool isn’t supported on erasure coded pools, and needs to be on a replicated pool. Results initially looked better, but the OSDs still exhausted their memory and caused host freezes, though after a longer time with more data successfully ingested.


I had intended to test RadosGW with the S3 API, but I decided against it. With two different failed tests, the chances for any test results that didn’t end with the cluster dying are pretty low.

The final cluster hardware setup, with three nodes mounted in a nice stack with shorter (and tidier) network cables. Blue cased Pi isn’t part of the cluster.


The Raspberry Pi 3 B+ doesn’t have enough RAM to run Ceph. While everything technically works, any attempts at sustained data transfer to or from the cluster fail. It seems like a Raspberry Pi, or other SBC, with 2GB+ of RAM would actually be stable, but still slow. The RAM issue is likely exacerbated by the 900kB/s random write rate the flash keys are capable of, but I don’t have faster flash keys or spare USB hard drives to test with.

Erasure coding seems to be better on RAM limited systems, and while it still failed, it always failed later, with more data successfully written. While it may have been more taxing on the Raspberry Pi’s limited CPU resources, these resources were typically in low contention, with usage averaging around 25% under maximum load across all 4 cores.

The release of the Raspberry Pi 4 B happened while I was writing this series of blog posts. I’d love to re-rerun these tests on three or four of the 4GB models with actual storage drives. The extra RAM should keep the OSDs from running out of memory and dying/taking down the whole system, and the USB3.0 interconnect means that storage and network access will be considerably faster. They might be good enough to run a small, yet stable cluster, and I look forward to testing on them soon.

Raspberry Pi Ceph Cluster – Testing Part 1

It’s time to run some tests on the Raspberry Pi Ceph cluster I built. I’m not sure if it’ll be stable enough to actually test, but I’d like to find out and try to tune things if needed.

Pool Creation:

I want to test both standard replicated pools, and Ceph’s newer erasure coded pools. I configured Ceph’s replicated pools with 2 replicas. The erasure coded pools are more like RAID in that there aren’t N replicas spread across the cluster, but rather data is split into chunks and distributed, checksummed, and then the data and checksums are spread across the pool. I configured the erasure coded pool with data split into 2 with an additional coding chunk. Practically, this means that they should tolerate the same failures, but with 1.5x the overhead instead of 2x the overhead.

I created one 16GiB pool of each type to test. Why not always use erasure coded pools? They’re more computationally complex, which might be bad on compute constrained devices such as the Raspberry Pi. They also don’t support the full set of operations. For example, RBD can’t completely reside on an erasure coded pool. There is a workaround, the metadata resides on a replicated pool with the data on the erasure coded pool.

Baseline Tests:

To get a basic idea for what performance level I could expect from the hardware underlying Ceph, I ran network and storage performance tests.

I tested network throughput with iperf between each Raspberry Pi 3 B+ and an external computer connected over Gigabit Ethernet. Each got around 250Mb/s, which is reasonable for a gigabit chip connected via USB2.0. For comparison, a Raspberry Pi 3 B (not the plus version) with Fast Ethernet tested around 95Mb/s. As a control, I also tested the same iperf client against another computer connected over full gigabit at 950Mb/s.

Disk throughput was tested using dd for sequential reads and writes and iometer for random reads and writes against a flash key with an XFS filesystem. XFS used to be the recommended filesystem for Ceph until Ceph released BlueStore, and it’s still used for BlueStore’s metadata storage partition. The 32GB flash keys performed at 32.7MB/s sequential read and 16.5 MB/s sequential write. Random read and write with 16kiB operations yielded 15MB/s and 0.9MB/s (that’s 900kB/s) respectively using sysbench’s fileio module.

RBD Tests:

RADOS Block Device (RBD) is a block storage service, so you can run your own filesystem but have Ceph’s replication protect the data as well as spread access over multiple drives for increased performance. Ceph currently doesn’t support using pure erasure coded pools for RBD (or CephFS), instead the data is stored in the erasure coded pool and the metedata in a replicated pool. Partial writes also need to be enabled per-pool, as per the docs.

Once the pools were created, I mounted each and started running tests. The first thing I wanted to test was just sequential read and write of data. To do this, I made an XFS filesystem with the defaults and mounted it on a test client (Arch Linux, kernel 5.1.14, quad core, 16GiB RAM, 1xGbE, Ceph 13.2.1) and wrote a file using dd if=/dev/zero of=/mnt/test/test.dd bs=4M count=3072 iflag=direct. Initially, writes to the replicated rbd image looked decent, averaging a whopping 6.4MB/s. And then the VM suddenly got really cranky.

The Ceph manager dashboard’s health status. Not what I’d been hoping for during a performance test.

One host, rpi-node2 had seemingly dropped from the network. Ceph went into recovery mode to keep my precious zeroes intact, and IO basically ground to a halt as the cluster recovered at a blazing 1.3MiB/s. I couldn’t hit the node with SSH, so I power-cycled it. It came back up, Ceph realized that the OSD was back and cut short the re-balancing of the cluster. I decided to run the write test again, deleted the test file and ran fstrim -v /mnt/test, which tells Ceph that the blocks can be freed, so it frees up that space on the OSDs so I could re-run fresh.

The second test ended similarly to the first, with dd writing happily at 6.1MB/s until rpi-node3 stopped responding (including to SSH) at 3.9GB written. This time I stopped dd immediately and waited for the node to come back, which is did after almost two minutes. I checked the logs and saw that the system was running out of memory and the ceph-osd process was getting OOM killed. I also noticed that both nodes that had failed were the ones running the active ceph-mgr instance serving the dashboards.

I ran the test again, this time generating a 1GiB file instead and confirmed that it was the node with the ceph-mgs instance running out of memory. I also let the write finish, testing how well Ceph ran in the degraded state. At 2.8MB/s, no performance records were being set, but the cluster was still ingesting data with one node missing.

I have two options, the first is to move the ceph-mgr daemon to another device, but as I wanted the cluster to be self-contained to three nodes, so I opted for the second option. Option two is to lower the memory usage of the ceph-osd daemon. I looked at the documentation for the BlueStore config, and saw that the default cache size is 1 GiB, or as much RAM as the Pi has. That just won’t do, so I added bluestore_cache_size = 536870912 and bluestore_cache_kv_max = 268435456 to my ceph.conf file and restarted the OSDs. This means that BlueStore will use at most 512MiB of RAM for its caches with only 256MiB maximum for the RocksDB metadata cache.

I reran the 1GiB file test and had 3.5M B/s write speed and no OSDs getting killed. With the 16GiB file, writes averaged 3MB/s, but RAM usage at the halfway mark eventually got the OSD running on the same node as the active ceph-mgr killed. Again, the cluster survived, just in a less happy state until the killed OSD daemon restarted. I disabled the dashboard, and while this helped RAM usage, the ceph-osd daemon was still getting killed. I further dropped the BlueStore cache to 256MiB and 128MiB for the metadata store. This time I locked two of the three nodes up hard and needed to power-cycle them.

During the 1GiB file tests, this was the unhappiest the cluster got. The PGs were simply being slow replicating and caught up quickly once the write load subsided.

With the replicated testing being close to a complete failure, I moved on to testing the erasure coded pool. I expected them to be worse due the the increased amount of compute resources needed. I was wrong. I was able to successfully write the test file, and the worst I got was an OSD being slow and not responding to the the monitor fast enough and then recovering a few seconds later. Sequential writes averaged 5.7MB/s and sequential read was an average of 6.1MB/s, but I still had two nodes go down at different times. It seems that erasure coded pools perform slightly better, but can still cause system memory exhaustion.

One thing to note is that even with three nodes, Ceph never lost any data and was still accepting write, just very slowly. I hadn’t intended to test Ceph’s resiliency, as that has already been well tested, but it was nice to see that it kept serving reads and writes.

At this point, I don’t think that the 1GiB of RAM is enough to run RBD properly. Sequential writes looked to be around 6MB/s when the cluster hadn’t lost any OSDs or whole hosts. I never attempted to test random access, due to the issues with sequential reads and writes.

CephFS and Rados Gateway:

With RBD being a bit of a bust, I wanted to see if CephFS and Rados Gateway performed better. As this post is getting long, CephFS and RadosGW results are in a second post, along with a conclusion.


Raspberry Pi Ceph Cluster – Setup

I’ve used Ceph several times in projects where I needed object storage or high resiliency. Other than using VMs, it’s never something I can easily run on at home to test, especially not on bare-metal. Circa 2013, I tried to get it running on a handful of Raspberry Pi Bs that I had, but they had far too little RAM to compile Ceph directly on, and I couldn’t get Ceph compiling reliably on a cross-compiler toolchain. So I moved on with life. Recently, I’ve been looking at Ceph as a more resilient (and expensive) replacement for my huge file server. I found that it’s now packaged on Arch Linux ARM’s repos. Perfect!

Having that hurdle out of the way, I decided to get a 3 node Ceph cluster running on some Raspberry Pis.

The Hardware:

Ceph Cluster Gear
All the gear needed to make a teeny Ceph cluster.

Hardware used in the cluster:

  • 3x Raspberry Pi 3 B+
  • 3x 32GB Sandisk Ultra microSD card (for operating system)
  • 3x 32GB Kingston DataTraveler USB key (for OSD)
  • 3x 2.5A microUSB power supplies
  • 3x Network cable
  • 1x Netgear GS108E Gigabit Switch and power supply

My plan is to have each node run all of the services needed for the cluster. This isn’t a recommended configuration, but I only have three Raspberry Pis to dedicate to this project. Additionally, I’m using a fourth Raspberry Pi as an automation/admin device, but it isn’t directly participating in the cluster.

Putting it all Together:

  • Create a master image to save to the three SD cards. I grabbed the latest image from the Arch Linux ARM project and followed their installation directions to get the first card set up.
  • Note: I used the Raspberry Pi 2 build, because the Ceph package for aarch64 on Arch Linux ARM is broken. I’m also using 13.2.1, because Arch’s version of Ceph is uncharacteristically almost a year old and I wasn’t going to try to compile 14.2.1 myself again.
  • Once the basic operating system image was installed, I put the card into the first Raspberry Pi and verified that SSH and networking came up as expected. It did, getting a static IP address from my already configured DHCP server.
  • I first tried to use Chef to configure the nodes once I got them running. Luckily Arch has a PKGBUILD for chef-client in the AUR. Unluckily, it’s not for aarch64. Luckily again, Chef is supported on aarch64 under Red Hat. I grabbed the PKGBUILD for x86_64, modified it to work with aarch64, built and installed the package.
  • I created a chef user, gave it sudo access, generated an SSH key on my chef-server, and copied it to the node.
  • At this point, I had done as much setup on the single node I had, so I copied the image onto the other two microSD cards, and put them into the other Pis.
  • Chef expects to be running on a system that isn’t Arch Linux. After some time trying to get it working, I decided that I’d spent enough time trying to get it working.
  • With Chef a bust, I moved on to Ansible and re-imaged the cards to start fresh.
  • ceph-ansible initially worked better, due to Ansible being supported on Arch Linux, but the playbook doesn’t actually support Arch. I needed to make some modifications to get the playbook to run.
  • With some basic configuration of the playbook ready to go, I got the mons up and running pretty easily. But osd creation failed on permissions issues. Something in the playbook was expecting a different configuration that Arch Linux uses. Adding the ceph user to the “disk” and “storage” groups partially fixed the permissions issues, but osd creation was still failing. Ugh.
Ceph cluster in operation. The rightmost blue Raspberry Pi is being used as an admin/automation server and isn’t part of the actual Ceph cluster.

Time for Manual Installation:

While part of my goal was to try some of the Ceph automation, chef’s and ceph-ansible’s lack of support for Arch Linux meant that I wasn’t really accomplishing my main goal, which was to get a small Raspberry Pi cluster up and running.

So I re-imaged the cards, used an Ansible bootstrap playbook that I wrote and referred to Ceph’s great manual deployment documentation. Why manually deploy it when ceph-deploy exists? Because ceph-deploy doesn’t support Arch Linux.


MONs or monitors stores the cluster map, which is used to determine where to store data to meet the reliability requirements. They also do general health monitoring of the cluster.

After following all of the steps in the Monitor Bootstrapping section, I had the first part of a working cluster, three monitors. Yay!

One difference from the official docs, in order to get the mons starting on boot, I needed to run systemctl enable ceph-mon.target in addition to the ceph-mon daemons, otherwise systemctl listed their status as Active: inactive (dead) and they didn’t start.


The next step was to get ceph-mgr running on the nodes with mons on them. The managers are used to provide a interfaces for external monitoring, as well as a web dashboard. While Ceph has documentation for this, I found Red Hat’s documentation more straight forward.

In order to enable the dashboard plugin, two things needed to be done on each node:

  • First, run ceph dashboard create-self-signed-cert to generate the self-signed SSL certificate used to secure the connection.
  • Then run ceph dashboard set-login-credentials username password, with the username and password credentials to create for the dashboard.
  • Running ceph mgr services then returned {"dashboard": "https://rpi-node1:8443/"}, confirming that things had worked correctly, and I could get the dashboard in my browser.


Now for the part of the cluster that will actually store my data, the OSD, which use the 32GB flash keys. If I wanted to add more flash keys, or maybe even hard drives, I could easily add more OSDs.

I followed the docs, adding one bluestore OSD per host on the flash key. One note, as I’d already tried to set the keys up using ceph-ansible, they did have GPT partition tables. I ran ceph-volume lvm zap /dev/sda on each host to fix this.

Additionally, I didn’t realize that the single ceph-volume command also sets up the systemd service, and I created my own. Now I had an OSD daemon without any storage in the cluster map. I followed Ceph’s documentation and removed the OSD, but now my OSDs IDs start at 1 instead of 0.


I plan on testing with CephFS, so the final step is to add the MetaData Server, or MDS, which stores metadata related to the filesystem. The ceph-mds daemon was enabled with systemctl enable ceph-mds@rpi-nodeN (N being the number of that node) and then systemctl enable ceph-mds.target so that the MDS is actually started.

Now What?

A healthy Raspberry Pi Ceph cluster, ready to go.

The cluster can’t store anything yet, because there aren’t any pools, radosgw hasn’t been set up and there aren’t any CephFS filesystems created.

If I had more Pis, I’d love to test thing with more complex CRUSH maps.

The next blog post will deal with testing of the cluster, which will likely include performance tuning. Raspberry Pis are very slow and resource constrained compared to the Xeon servers I’ve previously run Ceph on, so I expect things to go poorly with the default settings.


I split the results into two parts, the first includes test setup and RBD tests and the second will includes CephFS and radosgw tests as well as a conclusion.