On 07/08/2016 09:19 AM, Igor Fedotov wrote:
Varada,
thanks a lot for detailed overview.
Do I understand correctly that we are facing too large rocks db size
that doesn't fit into allocated bluefs space?
And now there is another question - why it becomes that large.
A couple of thoughts:
1) We need to see how well we've done with reducing the onode size.
2) Is anything leaking into the SST files that we don't want there?
3) How bad is the rocksdb write amp? You should be able to query
rocksdb or look at the periodic statistics in the log. We might be able
to improve it with tuning, but Allen said specifically this is an area
where Zetascale might help.
Mark
Will try to increase db size a bit later...
Thanks,
Igor
On 08.07.2016 16:52, Varada Kari wrote:
Default min allocation size on bluefs is 1MB.
Space consumption depends on the files created on the bluefs, if we have
more fnodes created on bluefs we would consume more space.
It depends on how many writes(new files) coming to bluefs from rocksdb
and space taken by the log file.
And for each modify/create/delete operations we log them in a file,
which gets compacted on each fsync from rocksdb upon meeting the
criteria to compact i.e; going beyond 16MB in size.
We try to allocate 4MB chunks for this file, to reserve some space for
the operation to continue.
In your case, we are failing to allocate 4MB for the compaction to
progress.
Can you try increasing the db partition size more and try it again?
Varada
On Friday 08 July 2016 06:44 PM, Igor Fedotov wrote:
Varada,
I'm running vstart.sh hence default block db and block wal sizes: 64 M &
128 M respectively.
Could you please elaborate why they go out of space - what information
is being accumulated there in my case? Actually I know a little about
this mechanics...
Thanks,
Igor
On 08.07.2016 14:29, Varada Kari wrote:
Igor,
How much is your db partition and wal partition size?
Seems you are out of space on db partition, block.db, if we go out of
space on block.db, we try to allocate block.wal.
Seems both are failing in your case and log compaction gone out of
space
for next 4MB chunk.
we shall discuss in the sync up today.
Varada
On Friday 08 July 2016 04:09 PM, Igor Fedotov wrote:
Hi All,
as I mentioned during yesterday's bluestore syncup I observed another
issue with bluestore during random write.
Here is the backtrace:
0> 2016-07-07 17:05:10.520543 7ff393dbf700 -1
os/bluestore/BlueFS.cc: In function 'int BlueFS::_allocate(unsigned
int,
uint64_t, std::vector<bluefs_extent_t>*)' thread 7ff393dbf700 time
2016-07-07 17:05:10.507412
os/bluestore/BlueFS.cc: 1398: FAILED assert(0 == "allocate
failed... wtf")
ceph version 11.0.0-289-g173e5a6
(173e5a6d85f624a714c0029db6f828cb1968cf3d)
1: (ceph::__ceph_assert_fail(char const*, char const*, int, char
const*)+0x82) [0x7ff3a9833452]
2: (BlueFS::_allocate(unsigned int, unsigned long,
std::vector<bluefs_extent_t, std::allocator<bluefs_extent_t>
>*)+0x760)
[0x7ff3a95186e0]
3: (BlueFS::_compact_log()+0xd9b) [0x7ff3a951ba8b]
4: (BlueFS::_maybe_compact_log()+0x2a0) [0x7ff3a951c510]
5: (BlueFS::sync_metadata()+0x20f) [0x7ff3a951d77f]
6: (BlueRocksDirectory::Fsync()+0xd) [0x7ff3a95300dd]
7: (rocksdb::DBImpl::WriteImpl(rocksdb::WriteOptions const&,
rocksdb::WriteBatch*, rocksdb::WriteCallback*, unsigned long*,
unsigned
long, bool)+0x14e1) [0x7ff3a95ea281]
8: (rocksdb::DBImpl::Write(rocksdb::WriteOptions const&,
rocksdb::WriteBatch*)+0x1b) [0x7ff3a95eae1b]
9:
(RocksDBStore::submit_transaction_sync(std::shared_ptr<KeyValueDB::TransactionImpl>)+0x6b)
[0x7ff3a95b7d7b]
10: (BlueStore::_kv_sync_thread()+0x167b) [0x7ff3a942aefb]
11: (BlueStore::KVSyncThread::entry()+0xd) [0x7ff3a944af5d]
12: (()+0x80a5) [0x7ff3a7a170a5]
13: (clone()+0x6d) [0x7ff3a58f9cfd]
NOTE: a copy of the executable, or `objdump -rdS <executable>` is
needed to interpret this.
Looks like bitmap allocator returns a failure at some moment.
My environment:
* Ceph cluster run via vstart.sh.
*rbd image created via:
./rbd create --size 1024 -c ceph.conf --image-feature layering
fio_test
./rbd map -c ceph.conf fio_test
* fio script is as follows:
[global]
[rbd_iodepth32]
ioengine=libaio
iodepth=32
filename=/dev/rbd12
size=1m
io_size=8192m
bs=128k
rw=randwrite
numjobs=3
Bug is easily reproducible with the script when bluestore allocator is
set to bitmap (by default). I was unable to reproduce the issue with
stupid allocator hence I suppose it's rather bitmap allocator internal
issue. Maybe some leaks as it occurs rather by the end of the FIO
script?
One should apply bluestore patch I posted yesterday prior to
reproduce
this issue as one can hit another bug otherwise.
Thanks,
Igor
--
To unsubscribe from this list: send the line "unsubscribe
ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
PLEASE NOTE: The information contained in this electronic mail
message is intended only for the use of the designated recipient(s)
named above. If the reader of this message is not the intended
recipient, you are hereby notified that you have received this
message in error and that any review, dissemination, distribution,
or copying of this message is strictly prohibited. If you have
received this communication in error, please notify the sender by
telephone or e-mail (as shown above) immediately and destroy any and
all copies of this message in your possession (whether hard copies
or electronically stored copies).
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
