Prelimilary results:
- CPU usage drops significantly. For example, in T1Q8 128K write test against
stub_uring_osd with 10G network and Athlon X4 860k CPU it drops from 100% to 30%
- Latency becomes slightly worse. In T1Q1 4K write test in the same environment
latency increases from 56 to 63 us.
- Small write throughput also becomes slightly worse. In T1Q128 4K write test
against stub iops decreases from 138k to ~110k (unstable, fluctuates 100k..120k).
Note that this is without io_uring, of course.
- Slightly reduce journaling write amplification (requires no_same_sector_overwrites=false)
- Fix listen_backlog (it was 0) because it could more than halve OSD socket send speed
- Support IPv6 OSD addresses
- Do not try to initialize client in simple-offsets
- Fix OSDs sometimes marking PGs incomplete instead of trying to connect with peers
- Allow to configure OSD placement in node_placement
- Allow to run with 4k sector size block devices. Natural, but it was forbidden
Slightly reduces WA. For example, in 4K T1Q128 replicated randwrite tests
WA is reduced from ~3.6 to ~3.1, in T1Q64 from ~3.8 to ~3.4.
Only effective without no_same_sector_overwrites.
- Implement a storage plugin for Proxmox. Now you can use Vitastor with Proxmox!
- Implement `vitastor-cli df` (pool space usage statistics) command
- Add glob pattern support for `vitastor-cli ls`
- Fix several bugs in other CLI commands (resize, create --parent, modify --readonly)
- Use 512 byte logical block size in QEMU driver by default (and thus don't require to set it in QEMU options)
New features:
- Build Vitastor driver as part of QEMU
- Implement renaming images in CLI (vitastor-cli modify --rename)
- Add vitastor-cli alloc-osd and simple-offsets commands and use them in make-osd,
thus removing the dependency on etcdctl
- Make monitor remove stale deleted inode statistics from etcd automatically
- Implement OSD address selection from a subnet, thus removing the need to specify
OSD addresses in startup scripts explicitly
Bug fixes:
- Fix client failover in case of etcd shutdown or crash (make client survive etcd failures)
- Stick to the last live etcd in OSD and mon to prevent random failures when one of etcds is down
- Fix incorrect copying of data from journal to the data device which could lead to data corruption
- Prefer local etcd IPs in OSD
- Remove the total PG count restriction in optimize_change which was sometimes leading
to inability to redistribute PGs over OSDs
- Fix error response parsing on a failed pg state report
- Fix slow linear writes with RDMA by changing default buffer settings
- Fix possible 'TypeError' in openstack nova when using Vitastor cinder driver
- Fix bugs in vitastor-cli create, ls, rm, modify commands
Patch changes:
- Add a patch for libvirt 7.6
- Add patches for QEMU 6.0 and 6.1
- Fix config file path XML location parsing in libvirt patches
- Replace _ with - in QEMU options
- Fix possible 'TypeError' in openstack nova when using Vitastor cinder driver
- Fix possible crashes of QEMU block driver in case of incorrect options
129K to leave extra space for the header
The problem with 8x 1M buffers is that the following happens with,
for example, 2 OSDs and 4M T1Q1 write:
- Server posts 8 receives
- Client posts 8 sends
- WRs are processed by the RDMA stack, but the OSD doesn't have the time
to handle them and doesn't refill buffers
- Client posts 1 more send
- RNR retransmission happens and performance drops to zero
Overall it seems that RDMA support should be reworked to use real 'RDMA'
operations i.e. operations writing into remote memory. This has an
additional advantage of avoiding a copy at the receive side of the OSD.
- Implement CLI commands for listing, viewing I/O statistics, creating,
snapshotting, cloning, resizing and modifying images. All these operations
are covered by 3 commands: ls, create, modify
- Implement an important fix to prior OSD set tracking for PGs. The previous
version had an issue which could lead to data loss due to an OSD with older
copy of the data thinking it has the newest copy
- Fix I/O statistics aggregation in the monitor
- Several minor fixes for Cinder driver
- Fix QEMU driver to be compatible with QEMU 2.x > 2.0
- Fix stalls sometimes possible in configurations without immediate_commit due
to insufficient amount of automatic internal fsync operations
- Add `vita` alias for `vitastor-cli`
Required to prevent data loss due to activation of an OSD with older data
when PG OSD set change doesn't occur. I.e. fixes the simplest case:
- Run 2 OSDs with 1 PG
- Start writing into the PG
- Stop OSD 2
- Stop OSD 1
- Start OSD 2
After this change the PG will refuse to start after the last step.
- New command-line tool: vitastor-cli
- Implement layer (snapshot/clone) merge and delete
- Remove 'bool' from the C header
- Fix a very rare flusher stall
- More diagnostics now printed for slow ops in the log
- Basic support for OpenStack: Cinder driver, patches for Nova and libvirt
- Add missing "image" and "config_path" QEMU options
- Calculate aggregate per-pool statistics in monitor
- Implement writes with Check-And-Set semantics
- Add a C wrapper library with public header
From now on, reads will return the server-side object version numbers
and writes and deletes will have an additional "version" parameter
which, if set to a non-zero value, will be atomically compared with
the current version of the object plus 1 and the modification will
fail if it doesn't match.
This feature opens the road to correct online flattening of snapshot
layers and other interesting things.
For CentOS 7 it also requires newer rdma-core as CentOS 7's native version doesn't have
implicit ODP support. The updated version is already uploaded into the vitastor repo.
This is the simplest and, as usual, the best implementation :)
100% zero-copy implementation is also possible (see rdma-zerocopy branch),
but it requires to create A LOT of queues (~128 per client) to use QPN as a 'tag'
because of the lack of receive tags and the server may simply run out of queues.
Hardware limit is 262144 on Mellanox ConnectX-4 which amounts to only 2048
'connections' per host. And even with that amount of queues it's still less optimal
than the non-zerocopy one.
In fact, newest hardware like Mellanox ConnectX-5 does have Tag Matching
support, but it's still unsuitable for us because it doesn't support scatter/gather
(tm_caps.max_sge=1).
Basic naive implementation works, but it's highly non-optimal as
RNR retransmissions occur all the time. RDMA expects the receiver
to always have place for incoming WRs...
Rework client operation queue from a vector to a linked list.
This is required to rework continue_ops() as its current implementation
consumes ~25% of client process CPU.
Warning: upgrading from 0.5.x is currently not supported!
Please create an issue if you really need upgrade capability.
New features:
- Snapshots and Copy-on-Write clones
- Inode (image) names
- Inode I/O and space statistics
- Write throttling for smoothing random write workloads in SSD+HDD configurations
The new protocol is almost compatible - it has bitmaps, but also it has
a "bitmap_length" field. It's not hard to make 0.5-0.6 OSDs and clients
compatible, but for now I just assume nobody needs it.
If I'm wrong and anybody requests to upgrade their production 0.5.x system
to 0.6.x I'll fix it.
Each inode has: image name, parent inode number & pool, size and readonly flag
Snapshots are created by switching image name to a different inode number
while using the older inode as parent.
The bug reproduced if fio was temporarily stopped with SIGSTOP
during write test and then resumed after 10 seconds. In this case
"pings" were failed for all clients and fio process crashed with
'use-after-free' in keepalive_timer. It happened because it called
stop_client while having a live iterator to the map.
This fixes a 'double_alloc' assertion in the following case:
- big_write object #1 v1 to block #100
- big_write object #1 v2 to block #101
- big_write object #2 v1 to block #100
Previous implementation didn't respect write ordering and could lead
to corrupted data when restarting writes after an OSD outage
Also rework cluster_client queueing logic and add tests for it to verify the correct behaviour
Previously BS_OP_SYNC could take unfinished writes and add them into the journal before
they were actually completed. This was leading to crashes with the message
"BUG: Unexpected dirty_entry 2000000000001:9f2a0000 v3 unstable state during flush: 338"
- Write operations could be 'stabilized' and previous versions could be
purged from OSDs before the removal of version_override and following
reads could potentially hit different version in EC pools
- Object was marked clean after completing the delete during recovery, so
reads could in theory hit a deleted version and return nothing
Previously OSDs could commit deletes before writes during recovery or rebalance
in the "lazy fsync" (immediate_commit=off) mode which could result in lost objects
This happened because flusher_count became dynamic and fsync_batch() was comparing the number
of flushers currently ready to do an fsync with the maximum number of flushers. Also the number
wasn't rechecked on every loop which was also incorrect.
Now the interrupted_rebalance test passes even without IMMEDIATE_COMMIT=1.
This fixes the bug where OSDs were sometimes trying to report updated PG states
infinitely without luck when PGs transitioned from 'starting' to 'peering' too fast
Make syncs wait for all previous writes because it's the only way
to make sure that OSDs do not receive incomplete writes in LIST results
during peering when some writes are still in progress.
Also simplify blockstore submission queue logic.
Vitaliy Filippov
Vitaliy Filippov