The qa tests are not client I/O centric and mostly focus on triggering
recovery/backfills and monitor them for completion within a finite amount
of time. The same holds true for scrub operations.
Therefore, an mClock profile that optimizes background operations is a
better fit for qa related tests. The osd_mclock_profile is therefore
globally overriden to 'high_recovery_ops' profile for the Rados suite as
it fits the requirement.
Also, many standalone tests expect recovery and scrub operations to
complete within a finite time. To ensure this, the osd_mclock_profile
options is set to 'high_recovery_ops' as part of the run_osd() function
in ceph-helpers.sh.
A subset of standalone tests explicitly used 'high_recovery_ops' profile.
Since the profile is now set as part of run_osd(), the earlier overrides
are redundant and therefore removed from the tests.
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Modify the relevant documentation to reflect:
- change in the default mClock profile to 'balanced'
- new allocations for ops across mClock profiles
- change in the osd_max_backfills limit
- miscellaneous changes related to warnings.
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
The osd_mclock_max_sequential_bandwidth_ssd is changed to 1200 MiB/s as
a reasonable middle ground considering the broad range of SSD capabilities.
This allows the mClock's cost model to extract the SSDs capability
depending on the cost of the IO being performed.
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
The earlier limit of 3 was still aggressive enough to have an impact on
the client and other competing operations. Retain the current default
for mClock. This can be modified if necessary after setting the
osd_mclock_override_recovery_settings option.
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Let's use the middle profile as the default.
Modify the standalone tests accordingly.
Signed-off-by: Samuel Just <sjust@redhat.com>
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Now that recovery operations are split between background_recovery and
background_best_effort, rebalance qos params to avoid penalizing
background_recovery while idle.
Signed-off-by: Samuel Just <sjust@redhat.com>
Otherwise, these end up as PGOpItem and therefore as immediate:
class PGOpItem : public PGOpQueueable {
...
op_scheduler_class get_scheduler_class() const final {
auto type = op->get_req()->get_type();
if (type == CEPH_MSG_OSD_OP ||
type == CEPH_MSG_OSD_BACKOFF) {
return op_scheduler_class::client;
} else {
return op_scheduler_class::immediate;
}
}
...
};
This was probably causing a bunch of extra interference with client
ops.
Signed-off-by: Samuel Just <sjust@redhat.com>
Recovery operations on pgs/objects that have fewer than the configured
number of copies should be treated more urgently than operations on
pgs/objects that simply need to be moved to a new location.
Signed-off-by: Samuel Just <sjust@redhat.com>
is_qos_item() was only used in operator<< for OpSchedulerItem. However,
it's actually useful to see priority for mclock items since it affects
whether it goes into the immediate queues and, for some types, the
class. Unconditionally display both class_id and priority.
Signed-off-by: Samuel Just <sjust@redhat.com>
Fix an issue where an overridden mClock recovery setting (set prior to
an osd restart) could be lost after an osd restart.
For e.g., consider that prior to an osd restart, the option
'osd_max_backfill' was successfully set to a value different from the
mClock default. If the osd was restarted for some reason, the
boot-up sequence was incorrectly resetting the backfill value to the
mclock default within the async local/remote reservers. This fix
ensures that no change is made if the current overriden value is
different from the mClock default.
Modify an existing standalone test to verify that the local and remote
async reservers are updated to the desired number of backfills under
normal conditions and also across osd restarts.
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Client ops are sensitive to the recovery load and must be carefully
set for osds whose underlying device is HDD. Tests revealed that
recoveries with osd_max_backfills = 10 and osd_recovery_max_active_hdd = 5
were still aggressive and overwhelmed client ops. The built-in defaults
for mClock are now set to:
1) osd_recovery_max_active_hdd = 3
2) osd_recovery_max_active_ssd = 10
3) osd_max_backfills = 3
The above may be modified if necessary by setting
osd_mclock_override_recovery_settings option.
Fixes: https://tracker.ceph.com/issues/58529
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Previously, setting default configs from the configured profile was
split across:
- enable_mclock_profile_settings
- set_mclock_profile - sets mclock_profile class member
- set_*_allocations - updates client_allocs class member
- set_profile_config - sets profile based on client_allocs class member
This made tracing the effect of changing the profile pretty challenging
due passing state through class member variables.
Instead, define a simple profile_t with three constexpr values
corresponding to the three profiles and handle it all in a single
set_config_defaults_from_profile() method.
Signed-off-by: Samuel Just <sjust@redhat.com>
The mClock scheduler's cost model for HDDs/SSDs is modified and now
represents the cost of an IO in terms of bytes.
The cost parameters, namely, osd_mclock_cost_per_io_usec_[hdd|ssd]
and osd_mclock_cost_per_byte_usec_[hdd|ssd] which represent the cost
of an IO in secs are inaccurate and therefore removed.
The new model considers the following aspects of an osd to calculate
the cost of an IO:
- osd_mclock_max_capacity_iops_[hdd|ssd] (existing option)
The measured random write IOPS at 4 KiB block size. This is
measured during OSD boot-up using OSD bench tool.
- osd_mclock_max_sequential_bandwidth_[hdd|ssd] (new config option)
The maximum sequential bandwidth of of the underlying device.
For HDDs, 150 MiB/s is considered, and for SSDs 750 MiB/s is
considered in the cost calculation.
The following important changes are made to arrive at the overall
cost of an IO,
1. Represent QoS reservation and limit config parameter as proportion:
The reservation and limit parameters are now set in terms of a
proportion of the OSD's max IOPS capacity. The earlier representation
was in terms of IOPS per OSD shard which required the user to perform
calculations before setting the parameter. Representing the
reservation and limit in terms of proportions is much more intuitive
and simpler for a user.
2. Cost per IO Calculation:
Using the above config options, osd_bandwidth_cost_per_io for the osd is
calculated and set. It is the ratio of the max sequential bandwidth and
the max random write iops of the osd. It is a constant and represents the
base cost of an IO in terms of bytes. This is added to the actual size of
the IO(in bytes) to represent the overall cost of the IO operation.See
mClockScheduler::calc_scaled_cost().
3. Cost calculation in Bytes:
The settings for reservation and limit in terms a fraction of the OSD's
maximum IOPS capacity is converted to Bytes/sec before updating the
mClock server's ClientInfo structure. This is done for each OSD op shard
using osd_bandwidth_capacity_per_shard shown below:
(res|lim) = (IOPS proportion) * osd_bandwidth_capacity_per_shard
(Bytes/sec) (unitless) (bytes/sec)
The above result is updated within the mClock server's ClientInfo
structure for different op_scheduler_class operations. See
mClockScheduler::ClientRegistry::update_from_config().
The overall cost of an IO operation (in secs) is finally determined
during the tag calculations performed in the mClock server. See
crimson::dmclock::RequestTag::tag_calc() for more details.
4. Profile Allocations:
Optimize mClock profile allocations due to the change in the cost model
and lower recovery cost.
5. Modify standalone tests to reflect the change in the QoS config
parameter representation of reservation and limit options.
Fixes: https://tracker.ceph.com/issues/58529
Fixes: https://tracker.ceph.com/issues/59080
Signed-off-by: Samuel Just <sjust@redhat.com>
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Previously, we used a static value of osd_recovery_cost (20M
by default) for PGRecovery. For pools with relatively small
objects, this causes mclock to backfill very very slowly as
20M massively overestimates the amount of IO each recovery
queue operation requires. Instead, add a cost_per_object
parameter to OSDService::awaiting_throttle and set it to the
average object size in the PG being queued.
Fixes: https://tracker.ceph.com/issues/58606
Signed-off-by: Samuel Just <sjust@redhat.com>
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Previously, we always queued this with cost osd_recovery_cost which
defaults to 20M. With mclock, this caused these items to be delayed
heavily. Instead, base the cost on the operation queued.
Fixes: https://tracker.ceph.com/issues/58606
Signed-off-by: Samuel Just <sjust@redhat.com>
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
See included comments -- previous values did not account for object
size. This causes problems for mclock which is much more strict
in how it interprets costs.
Fixes: https://tracker.ceph.com/issues/58607
Signed-off-by: Samuel Just <sjust@redhat.com>
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
See included comments -- previous values did not account for object
size. This causes problems for mclock which is much more strict
in how it interprets costs.
Fixes: https://tracker.ceph.com/issues/58529
Signed-off-by: Samuel Just <sjust@redhat.com>
Signed-off-by: Sridhar Seshasayee <sseshasa@redhat.com>
Add a procedure to doc/cephfs/troubleshooting.rst that explains how to
restore access to FileSystems that became inaccessible after
post-Nautilus upgrades. The procedure included here was written by Harry
G Coin, and merely lightly edited by me. I include him here as a
"co-author", but it should be noted that he did the heavy lifting on
this.
See the email thread here for more context:
https://lists.ceph.io/hyperkitty/list/ceph-users@ceph.io/message/HS5FD3QFR77NAKJ43M2T5ZC25UYXFLNW/
Co-authored-by: Harry G Coin <hgcoin@gmail.com>
Signed-off-by: Zac Dover <zac.dover@proton.me>
The previous `ceph osd tree` doesn't show pool-defined weight-sets as the above documentation suggests.
Signed-off-by: James Lakin <james@jameslakin.co.uk>
* refs/pull/51224/head:
doc: add a note for minimum compatible python version and supported distros
tools/cephfs/top/CMakeList.txt: check the minimum compatible python version for cephfs-top
Reviewed-by: Venky Shankar <vshankar@redhat.com>
rgw/cloud-transition: Handle cloud-tiered objects in a multisite environment
Reviewed-by: Casey Bodley <cbodley@redhat.com>
Reviewed-by: Matt Benjamin <mbenjamin@redhat.com>
