Troubleshooting OSDs

Before troubleshooting your OSDs, first check your monitors and network. If you execute ceph health or ceph -s on the command line and Ceph shows HEALTH_OK, it means that the monitors have a quorum. If you don’t have a monitor quorum or if there are errors with the monitor status, address the monitor issues first. Check your networks to ensure they are running properly, because networks may have a significant impact on OSD operation and performance. Look for dropped packets on the host side and CRC errors on the switch side.

Obtaining Data About OSDs

A good first step in troubleshooting your OSDs is to obtain topology information in addition to the information you collected while monitoring your OSDs (e.g., ceph osd tree).

Ceph Logs

If you haven’t changed the default path, you can find Ceph log files at /var/log/ceph:

ls /var/log/ceph

If you don’t see enough log detail you can change your logging level. See Logging and Debugging for details to ensure that Ceph performs adequately under high logging volume.

Admin Socket

Use the admin socket tool to retrieve runtime information. For details, list the sockets for your Ceph daemons:

ls /var/run/ceph

Then, execute the following, replacing {daemon-name} with an actual daemon (e.g., osd.0):

ceph daemon osd.0 help

Alternatively, you can specify a {socket-file} (e.g., something in /var/run/ceph):

ceph daemon {socket-file} help

The admin socket, among other things, allows you to:

  • List your configuration at runtime

  • Dump historic operations

  • Dump the operation priority queue state

  • Dump operations in flight

  • Dump perfcounters

Display Freespace

Filesystem issues may arise. To display your file system’s free space, execute df.

df -h

Execute df --help for additional usage.

I/O Statistics

Use iostat to identify I/O-related issues.

iostat -x

Diagnostic Messages

To retrieve diagnostic messages from the kernel, use dmesg with less, more, grep or tail. For example:

dmesg | grep scsi

Stopping w/out Rebalancing

Periodically, you may need to perform maintenance on a subset of your cluster, or resolve a problem that affects a failure domain (e.g., a rack). If you do not want CRUSH to automatically rebalance the cluster as you stop OSDs for maintenance, set the cluster to noout first:

ceph osd set noout

On Luminous or newer releases it is safer to set the flag only on affected OSDs. You can do this individually

ceph osd add-noout osd.0
ceph osd rm-noout  osd.0

Or an entire CRUSH bucket at a time. Say you’re going to take down prod-ceph-data1701 to add RAM

ceph osd set-group noout prod-ceph-data1701

Once the flag is set you can begin stopping the OSDs within the failure domain that requires maintenance work.

stop ceph-osd id={num}

Note

Placement groups within the OSDs you stop will become degraded while you are addressing issues with within the failure domain.

Once you have completed your maintenance, restart the OSDs.

start ceph-osd id={num}

Finally, you must unset the cluster from noout.

ceph osd unset noout
ceph osd unset-group noout prod-ceph-data1701

OSD Not Running

Under normal circumstances, simply restarting the ceph-osd daemon will allow it to rejoin the cluster and recover.

An OSD Won’t Start

If you start your cluster and an OSD won’t start, check the following:

  • Configuration File: If you were not able to get OSDs running from a new installation, check your configuration file to ensure it conforms (e.g., host not hostname, etc.).

  • Check Paths: Check the paths in your configuration, and the actual paths themselves for data and metadata (journals, WAL, DB). If you separate the OSD data from the metadata and there are errors in your configuration file or in the actual mounts, you may have trouble starting OSDs. If you want to store the metadata on a separate block device, you should partition or LVM your drive and assign one partition per OSD.

  • Check Max Threadcount: If you have a node with a lot of OSDs, you may be hitting the default maximum number of threads (e.g., usually 32k), especially during recovery. You can increase the number of threads using sysctl to see if increasing the maximum number of threads to the maximum possible number of threads allowed (i.e., 4194303) will help. For example:

    sysctl -w kernel.pid_max=4194303
    

    If increasing the maximum thread count resolves the issue, you can make it permanent by including a kernel.pid_max setting in a file under /etc/sysctl.d or within the master /etc/sysctl.conf file. For example:

    kernel.pid_max = 4194303
    
  • Check ``nf_conntrack``: This connection tracking and limiting system is the bane of many production Ceph clusters, and can be insidious in that everything is fine at first. As cluster topology and client workload grow, mysterious and intermittent connection failures and performance glitches manifest, becoming worse over time and at certain times of day. Check syslog history for table fillage events. You can mitigate this bother by raising nf_conntrack_max to a much higher value via sysctl. Be sure to raise nf_conntrack_buckets accordingly to nf_conntrack_max / 4, which may require action outside of sysctl e.g. "echo 131072 > /sys/module/nf_conntrack/parameters/hashsize More interdictive but fussier is to blacklist the associated kernel modules to disable processing altogether. This is fragile in that the modules vary among kernel versions, as does the order in which they must be listed. Even when blacklisted there are situations in which iptables or docker may activate connection tracking anyway, so a “set and forget” strategy for the tunables is advised. On modern systems this will not consume appreciable resources.

  • Kernel Version: Identify the kernel version and distribution you are using. Ceph uses some third party tools by default, which may be buggy or may conflict with certain distributions and/or kernel versions (e.g., Google gperftools and TCMalloc). Check the OS recommendations and the release notes for each Ceph version to ensure you have addressed any issues related to your kernel.

  • Segment Fault: If there is a segment fault, increase log levels and start the problematic daemon(s) again. If segment faults recur, search the Ceph bug tracker https://tracker.ceph/com/projects/ceph and the dev and ceph-users mailing list archives https://ceph.io/resources. If this is truly a new and unique failure, post to the dev email list and provide the specific Ceph release being run, ceph.conf (with secrets XXX’d out), your monitor status output and excerpts from your log file(s).

An OSD Failed

When a ceph-osd process dies, surviving ceph-osd daemons will report to the mons that it appears down, which will in turn surface the new status via the ceph health command:

ceph health
HEALTH_WARN 1/3 in osds are down

Specifically, you will get a warning whenever there are OSDs marked in and down. You can identify which are down with:

ceph health detail
HEALTH_WARN 1/3 in osds are down
osd.0 is down since epoch 23, last address 192.168.106.220:6800/11080

or

ceph osd tree down

If there is a drive failure or other fault preventing ceph-osd from functioning or restarting, an error message should be present in its log file under /var/log/ceph.

If the daemon stopped because of a heartbeat failure or suicide timeout, the underlying drive or filesystem may be unresponsive. Check dmesg and syslog output for drive or other kernel errors. You may need to specify something like dmesg -T to get timestamps, otherwise it’s easy to mistake old errors for new.

If the problem is a software error (failed assertion or other unexpected error), search the archives and tracker as above, and report it to the ceph-devel email list if there’s no clear fix or existing bug.

No Free Drive Space

Ceph prevents you from writing to a full OSD so that you don’t lose data. In an operational cluster, you should receive a warning when your cluster’s OSDs and pools approach the full ratio. The mon osd full ratio defaults to 0.95, or 95% of capacity before it stops clients from writing data. The mon osd backfillfull ratio defaults to 0.90, or 90 % of capacity above which backfills will not start. The OSD nearfull ratio defaults to 0.85, or 85% of capacity when it generates a health warning.

Note that individual OSDs within a cluster will vary in how much data Ceph allocates to them. This utilization can be displayed for each OSD with

ceph osd df

Overall cluster / pool fullness can be checked with

ceph df

Pay close attention to the most full OSDs, not the percentage of raw space used as reported by ceph df. It only takes one outlier OSD filling up to fail writes to its pool. The space available to each pool as reported by ceph df considers the ratio settings relative to the most full OSD that is part of a given pool. The distribution can be flattened by progressively moving data from overfull or to underfull OSDs using the reweight-by-utilization command. With Ceph releases beginning with later revisions of Luminous one can also exploit the ceph-mgr balancer module to perform this task automatically and rather effectively.

The ratios can be adjusted:

ceph osd set-nearfull-ratio <float[0.0-1.0]>
ceph osd set-full-ratio <float[0.0-1.0]>
ceph osd set-backfillfull-ratio <float[0.0-1.0]>

Full cluster issues can arise when an OSD fails either as a test or organically within small and/or very full or unbalanced cluster. When an OSD or node holds an outsize percentage of the cluster’s data, the nearfull and full ratios may be exceeded as a result of component failures or even natural growth. If you are testing how Ceph reacts to OSD failures on a small cluster, you should leave ample free disk space and consider temporarily lowering the OSD full ratio, OSD backfillfull ratio and OSD nearfull ratio

Full ceph-osds will be reported by ceph health:

ceph health
HEALTH_WARN 1 nearfull osd(s)

Or:

ceph health detail
HEALTH_ERR 1 full osd(s); 1 backfillfull osd(s); 1 nearfull osd(s)
osd.3 is full at 97%
osd.4 is backfill full at 91%
osd.2 is near full at 87%

The best way to deal with a full cluster is to add capacity via new OSDs, enabling the cluster to redistribute data to newly available storage.

If you cannot start a legacy Filestore OSD because it is full, you may reclaim some space deleting a few placement group directories in the full OSD.

Important

If you choose to delete a placement group directory on a full OSD, DO NOT delete the same placement group directory on another full OSD, or YOU WILL LOSE DATA. You MUST maintain at least one copy of your data on at least one OSD. This is a rare and extreme intervention, and is not to be undertaken lightly.

See Monitor Config Reference for additional details.

OSDs are Slow/Unresponsive

A commonly recurring issue involves slow or unresponsive OSDs. Ensure that you have eliminated other troubleshooting possibilities before delving into OSD performance issues. For example, ensure that your network(s) is working properly and your OSDs are running. Check to see if OSDs are throttling recovery traffic.

Tip

Newer versions of Ceph provide better recovery handling by preventing recovering OSDs from using up system resources so that up and in OSDs are not available or are otherwise slow.

Networking Issues

Ceph is a distributed storage system, so it relies upon networks for OSD peering and replication, recovery from faults, and periodic heartbeats. Networking issues can cause OSD latency and flapping OSDs. See Flapping OSDs for details.

Ensure that Ceph processes and Ceph-dependent processes are connected and/or listening.

netstat -a | grep ceph
netstat -l | grep ceph
sudo netstat -p | grep ceph

Check network statistics.

netstat -s

Drive Configuration

A SAS or SATA storage drive should only house one OSD; NVMe drives readily handle two or more. Read and write throughput can bottleneck if other processes share the drive, including journals / metadata, operating systems, Ceph monitors, syslog logs, other OSDs, and non-Ceph processes.

Ceph acknowledges writes after journaling, so fast SSDs are an attractive option to accelerate the response time–particularly when using the XFS or ext4 file systems for legacy Filestore OSDs. By contrast, the Btrfs file system can write and journal simultaneously. (Note, however, that we recommend against using Btrfs for production deployments.)

Note

Partitioning a drive does not change its total throughput or sequential read/write limits. Running a journal in a separate partition may help, but you should prefer a separate physical drive.

Bad Sectors / Fragmented Disk

Check your drives for bad blocks, fragmentation, and other errors that can cause performance to drop substantially. Invaluable tools include dmesg, syslog logs, and smartctl (from the smartmontools package).

Co-resident Monitors/OSDs

Monitors are relatively lightweight processes, but they issue lots of fsync() calls, which can interfere with other workloads, particularly if monitors run on the same drive as an OSD. Additionally, if you run monitors on the same host as OSDs, you may incur performance issues related to:

  • Running an older kernel (pre-3.0)

  • Running a kernel with no syncfs(2) syscall.

In these cases, multiple OSDs running on the same host can drag each other down by doing lots of commits. That often leads to the bursty writes.

Co-resident Processes

Spinning up co-resident processes (convergence) such as a cloud-based solution, virtual machines and other applications that write data to Ceph while operating on the same hardware as OSDs can introduce significant OSD latency. Generally, we recommend optimizing hosts for use with Ceph and using other hosts for other processes. The practice of separating Ceph operations from other applications may help improve performance and may streamline troubleshooting and maintenance.

Logging Levels

If you turned logging levels up to track an issue and then forgot to turn logging levels back down, the OSD may be putting a lot of logs onto the disk. If you intend to keep logging levels high, you may consider mounting a drive to the default path for logging (i.e., /var/log/ceph/$cluster-$name.log).

Recovery Throttling

Depending upon your configuration, Ceph may reduce recovery rates to maintain performance or it may increase recovery rates to the point that recovery impacts OSD performance. Check to see if the OSD is recovering.

Kernel Version

Check the kernel version you are running. Older kernels may not receive new backports that Ceph depends upon for better performance.

Kernel Issues with SyncFS

Try running one OSD per host to see if performance improves. Old kernels might not have a recent enough version of glibc to support syncfs(2).

Filesystem Issues

Currently, we recommend deploying clusters with the BlueStore back end. When running a pre-Luminous release or if you have a specific reason to deploy OSDs with the previous Filestore backend, we recommend XFS.

We recommend against using Btrfs or ext4. The Btrfs filesystem has many attractive features, but bugs may lead to performance issues and spurious ENOSPC errors. We do not recommend ext4 for Filestore OSDs because xattr limitations break support for long object names, which are needed for RGW.

For more information, see Filesystem Recommendations.

Insufficient RAM

We recommend a minimum of 4GB of RAM per OSD daemon and suggest rounding up from 6-8GB. You may notice that during normal operations, ceph-osd processes only use a fraction of that amount. Unused RAM makes it tempting to use the excess RAM for co-resident applications or to skimp on each node’s memory capacity. However, when OSDs experience recovery their memory utilization spikes. If there is insufficient RAM available, OSD performance will slow considerably and the daemons may even crash or be killed by the Linux OOM Killer.

Blocked Requests or Slow Requests

If a ceph-osd daemon is slow to respond to a request, messages will be logged noting ops that are taking too long. The warning threshold defaults to 30 seconds and is configurable via the osd op complaint time setting. When this happens, the cluster log will receive messages.

Legacy versions of Ceph complain about old requests:

osd.0 192.168.106.220:6800/18813 312 : [WRN] old request osd_op(client.5099.0:790 fatty_26485_object789 [write 0~4096] 2.5e54f643) v4 received at 2012-03-06 15:42:56.054801 currently waiting for sub ops

New versions of Ceph complain about slow requests:

{date} {osd.num} [WRN] 1 slow requests, 1 included below; oldest blocked for > 30.005692 secs
{date} {osd.num}  [WRN] slow request 30.005692 seconds old, received at {date-time}: osd_op(client.4240.0:8 benchmark_data_ceph-1_39426_object7 [write 0~4194304] 0.69848840) v4 currently waiting for subops from [610]

Possible causes include:

  • A failing drive (check dmesg output)

  • A bug in the kernel file system (check dmesg output)

  • An overloaded cluster (check system load, iostat, etc.)

  • A bug in the ceph-osd daemon.

Possible solutions:

  • Remove VMs from Ceph hosts

  • Upgrade kernel

  • Upgrade Ceph

  • Restart OSDs

  • Replace failed or failing components

Debugging Slow Requests

If you run ceph daemon osd.<id> dump_historic_ops or ceph daemon osd.<id> dump_ops_in_flight, you will see a set of operations and a list of events each operation went through. These are briefly described below.

Events from the Messenger layer:

  • header_read: When the messenger first started reading the message off the wire.

  • throttled: When the messenger tried to acquire memory throttle space to read the message into memory.

  • all_read: When the messenger finished reading the message off the wire.

  • dispatched: When the messenger gave the message to the OSD.

  • initiated: This is identical to header_read. The existence of both is a historical oddity.

Events from the OSD as it processes ops:

  • queued_for_pg: The op has been put into the queue for processing by its PG.

  • reached_pg: The PG has started doing the op.

  • waiting for \*: The op is waiting for some other work to complete before it can proceed (e.g. a new OSDMap; for its object target to scrub; for the PG to finish peering; all as specified in the message).

  • started: The op has been accepted as something the OSD should do and is now being performed.

  • waiting for subops from: The op has been sent to replica OSDs.

Events from `Filestore`:

  • commit_queued_for_journal_write: The op has been given to the FileStore.

  • write_thread_in_journal_buffer: The op is in the journal’s buffer and waiting to be persisted (as the next disk write).

  • journaled_completion_queued: The op was journaled to disk and its callback queued for invocation.

Events from the OSD after data has been given to underlying storage:

  • op_commit: The op has been committed (i.e. written to journal) by the primary OSD.

  • op_applied: The op has been write()’en to the backing FS (i.e. applied in memory but not flushed out to disk) on the primary.

  • sub_op_applied: op_applied, but for a replica’s “subop”.

  • sub_op_committed: op_commit, but for a replica’s subop (only for EC pools).

  • sub_op_commit_rec/sub_op_apply_rec from <X>: The primary marks this when it hears about the above, but for a particular replica (i.e. <X>).

  • commit_sent: We sent a reply back to the client (or primary OSD, for sub ops).

Many of these events are seemingly redundant, but cross important boundaries in the internal code (such as passing data across locks into new threads).

Flapping OSDs

When OSDs peer and check heartbeats, they use the cluster (back-end) network when it’s available. See Monitor/OSD Interaction for details.

We have tradtionally recommended separate public (front-end) and private (cluster / back-end / replication) networks:

  1. Segregation of heartbeat and replication / recovery traffic (private) from client and OSD <-> mon traffic (public). This helps keep one from DoS-ing the other, which could in turn result in a cascading failure.

  2. Additional throughput for both public and private traffic.

When common networking technloogies were 100Mb/s and 1Gb/s, this separation was often critical. With today’s 10Gb/s, 40Gb/s, and 25/50/100Gb/s networks, the above capacity concerns are often diminished or even obviated. For example, if your OSD nodes have two network ports, dedicating one to the public and the other to the private network means no path redundancy. This degrades your ability to weather network maintenance and failures without significant cluster or client impact. Consider instead using both links for just a public network: with bonding (LACP) or equal-cost routing (e.g. FRR) you reap the benefits of increased throughput headroom, fault tolerance, and reduced OSD flapping.

When a private network (or even a single host link) fails or degrades while the public network operates normally, OSDs may not handle this situation well. What happens is that OSDs use the public network to report each other down to the monitors, while marking themselves up. The monitors then send out, again on the public network, an updated cluster map with affected OSDs marked down. These OSDs reply to the monitors “I’m not dead yet!”, and the cycle repeats. We call this scenario ‘flapping`, and it can be difficult to isolate and remediate. With no private network, this irksome dynamic is avoided: OSDs are generally either up or down without flapping.

If something does cause OSDs to ‘flap’ (repeatedly getting marked down and then up again), you can force the monitors to halt the flapping by temporarily freezing their states:

ceph osd set noup      # prevent OSDs from getting marked up
ceph osd set nodown    # prevent OSDs from getting marked down

These flags are recorded in the osdmap:

ceph osd dump | grep flags
flags no-up,no-down

You can clear the flags with:

ceph osd unset noup
ceph osd unset nodown

Two other flags are supported, noin and noout, which prevent booting OSDs from being marked in (allocated data) or protect OSDs from eventually being marked out (regardless of what the current value for mon osd down out interval is).

Note

noup, noout, and nodown are temporary in the sense that once the flags are cleared, the action they were blocking should occur shortly after. The noin flag, on the other hand, prevents OSDs from being marked in on boot, and any daemons that started while the flag was set will remain that way.

Note

The causes and effects of flapping can be somewhat mitigated through careful adjustments to the mon_osd_down_out_subtree_limit, mon_osd_reporter_subtree_level, and mon_osd_min_down_reporters. Derivation of optimal settings depends on cluster size, topology, and the Ceph release in use. Their interactions are subtle and beyond the scope of this document.