I've been playing around with this over the last couple of days.
It doesn't seem to have changed much since Alex wrote his paper in 2007,
including the bit about the bitmap writes to bct file not being
instrumented!
A few thoughts:
Have you tried running the script that Alex wrote (possibly with a couple
of tweaks to allow for all your data files) to estimate the size of the BCT
file to check how his estimate compares with your actual.
If you increase a buffer size then it will take longer to process the
buffer. so if the "reliable message" is something to do with informing
sessions that the buffer is clear, or that a checkpoint is complete then a
bigger buffer ***MAY*** mean you have to wait more often and longer for
that message to arrive.
It seems that the bitmaps are triggered when CKPT triggers a checkpoint.
Without access to a database of your size with concurrent activity of your
rate I can't do any realistic testing, but maybe ALL checkpoints (including
object checkpoints) are triggering the CTWR writes - so maybe you get the
long waits when you get several sessions starting direct path tablescans at
about the same time
You could use oradebug to set event 10046 level 8 for the ctwr process to
see something of what it's doing. The only things I saw in 11g were waits
for reliable message with a 3 second timeout and some sort of read/write
heartbeat on the change tracking file once every minute. This heartbeat
doesn't seem to have anything to do with the _bct_health_check_interval
(which happens to be 60 seconds by default).
oradebug setospid {os process id for ctwr}
oradebug event 10046 trace name context forever, level 8
(you could also use "osdebug setorapid {v$process.pid for ctwr}" instead of
setospid.
As far as I could tell a session may be writing block ids into a private
CTWR buffer as it write redo to the log buffer, and then something triggers
a transfer from the private buffer to the public bct file buffer, and that
transfer MAY include a "deduplication" of the block ids in the private
buffer. So if you have a large number of changes in a single database block
you could get a lot of entries in a ctwr private buffer that have to be
collapsed into a single bit in the public buffer. This may be one thing
that causes peaks in processing time. There is also a
"_bct_initial_private_dba_buffer_size" that might have some impact on how
private buffers make sessions stall, the default is 50 (entries) and maybe
a larger size would help.
Have you tried running queries against x$kcrstat every few seconds when the
system is busy to see if that shows you any intersting results ?
I would chase Oracle to see if they can tell you anything about adjusting
ctwr through events 19774 and 19775.
Regards
Jonathan Lewis
On Sat, 24 Apr 2021 at 13:35, Lok P <loknath.73@xxxxxxxxx> wrote:
One of the latest suggestions we are getting is to set below additional
parameters along with _bct_buffer_allocation_max(which we already
increased to ~1GB). But what we are wondering is ,in below the
_bct_public_dba_buffer_size has been set to ~190MB which is smaller than
current value(i.e. ~256MB) will it really going to help or going to make
situation worse?
We are thinking to atleast set first two parameter from below as those are
online changes. Last one ( _high_priority_processes) will need DB reboot,
so we are thinking to apply afterwards when we can afford downtime. But we
are still checking if _bct_public_dba_buffer_size should be increased
higher from current ~256Mb rather getting decreased.
alter system set "_bct_public_dba_buffer_size"=200000000 scope=both
sid='*';
alter system set "_bct_health_check_interval"=0 scope=both sid='*';
alter system set "_high_priority_processes"=CTWR scope=both sid='*';
On Sat, Apr 24, 2021 at 5:33 PM Lok P <loknath.73@xxxxxxxxx> wrote:
Thank You Jonathan. Yes the BCT file size was down from 22GB to 9GB after
we reinitiated.
Total size of dba_data_files= 148TB,
Count of data file = 3834
Its ZDLRA RMAN backup in place which means we have only one L0(FULL) i.e.
the very first backup and rest all subsequent backups are L1(Incremental)
only.
Its a 4-Node RAC and database version is 11.2.0.4.
One doubt though , I am not very sure if its related , but it seems the
issue surfaced after we created one new big file tablespace. Actually we
were having all the normal tablespaces but for one of the tablespace we
reached the max number of data file per tablespace limit. So we created big
file tablespace, so not sure if it has any role to play here.
Another observation is , last time I remember, when we encountered the
issue in month of february , we were seeing lot of "bct buffer space" wait
but very minimal "reliable message" wait. And we were seeing those waits
("bct buffer space" ) experienced by DMLS continuously and we were stuck.
But this time we are seeing a lot of "reliable message waits" but minimal
"BCT buffer space" waits and this is intermittent. So this time when this
waits(reliable message) appears thing gets just hung and then after 30-45
minutes it recovers automatically.
And last time (in feb) as suggested by the oracle , we did increased the
_bct_buffer_allocation_max to 640MB and large_pool size to 2GB but that
didn't help us. The again we reached out to oracle and they suggested to
increase the size of _bct_public_dba_buffer_size from ~128MB to ~256MB and
reinitiated the BCT and this helped in mitigating the issue. But it
resurfaced after 2months again now(i.e. in April this time).
And this time we tried by simply reinitiating BCT multiple times as an
work around and the issue is mitigating immediately but is surfacing after
2days from reinitiating the BCT while the BCT file size remains <10GB. So
definitely its not just the BCT filesize causing this but something else.
And this time, Oracle insisted to again bump _bct_buffer_allocation_max
to ~1GB , but we already did it and it didn't fix the issue so far. So we
were thinking if just like last time, increasing size of
_bct_public_dba_buffer_size further (may be ~1GB) will fix this issue. But
we have not got nod from Oracle yet on this. Do you think we should
consider increasing _bct_public_dba_buffer_size ?
Regards
Lok
On Sat, Apr 24, 2021 at 3:20 PM Jonathan Lewis <jlewisoracle@xxxxxxxxx>
wrote:
Just finding time to look at this - it's something I'd need to research
since I don't have anything but the barest of old notes in my library.
I know I sent you an indication of how large the BCT file probably ought
to be for your size database, but I can't find that message at present and
I may have a correction to it that explains the 22GB / 9GB.
How big is the database - sum(bytes)/(2^9), count(*) from dba_data_files
How many nodes (4 ?)
How many backups are you keeping before obsoleting them.
What was the version of Oracle (I keep forgetting)
I think I said that you need to estimate 1 bit per 32KB of data file,
multiplied by 8 as the total number of bitmaps - then allow rounding errors.
I've just found a note on MOS which says you also need to multiply by
the number of threads - and I don't think I knew that (though it makes
sense because you don't want the the instances to have to tell each other
about changes they've made and you don't want CTWRs to collide while trying
to write the same bits of bitmap).
-----
The "checkpoint pattern" for waits for reliable message is one that
could be repeated here - process A calls process B to get something done
and waits on a reliable message, process B calls process C to perform an
action, takes an exclusive lock, then messages process A which tries to
take a shared lock; process B waits for process C to do its thing and
release its lock when process C completes, which allows process A to get
(and release) its lock.
What we need to find out is (a) is this what's happening then (b) who
are process B and C, then (c) why is it taking so long for process B to get
acknowledgement from process C that it's working on the required task. (If
process A is waiting on reliable message this means process B hasn't heard
from process C or process B is unable to gets its exclusive lock).
A first step would be to get the wait chain analysis from
g$active_session_history as soon as possible after one of these long waits
has happened.
I'll be reading your previous couple of emails over the weekend and
hunting around MOS to see what I can find. I am a little curious that you
mention that CTWR writes on a checkpoint - this doesn't seem to be
necessary (or a good idea) - but that may be explained in one of the
notes/bugs you've found.
A couple of speculative thoughts:
1) possibly the problem appears when the BCT file has to be extended by
multiple instances at the same time and they have to queue.
2) maybe the buffer space waits follow the reliable message wait because
a lot of change has happened in the interval and the bitmap changes have
been queued up so long that they keep filling the buffer and waiting for
space, so maybe increasing the publc buffer size again would reduce those
waits (while we try to find out why the reliable message waits are
happening).
Regards
Jonathan Lewis
On Fri, 16 Apr 2021 at 20:36, Lok P <loknath.73@xxxxxxxxx> wrote:
We have re-initiated the BCT by disabling and then enabling it again
and then we saw the BCT file size decreased from 22GB to ~9GB and the wait
events "bct buffer space" and "reliable message" went away, everyone was
happy. But again after 2 days we again started seeing the "reliable
message" wait event popping up along with a few "bct buffer space" waits.
We are thinking of reaching out to Oracle through SR to see what is causing
this issue. If it's anything related to our current ZDLRA backup strategy
which has only one L0 and all L1's or something else is causing this issue?
Or we really have to increase the BCT buffer(
_bct_public_dba_buffer_size ) till ~1GB(which is the max size for it).
On Sun, Apr 11, 2021 at 12:53 AM Lok P <loknath.73@xxxxxxxxx> wrote:
As I read it , when a checkpoint occurs the change tracking writer
i.e. CTWR will write the data from the CTWR DBA buffer to the BCT file, so
we are seeing this BCT Buffer space wait + reliable message waits means
the
check point is not finishing up and getting stuck. So does it mean that
when we issue COMMIT during that point itself the changed data
is struggling to get written to CTWR buffer because of its so small space?
But again if it's because of the small size of CTWR buffer being small
then
how come initiating BCT file fixing such an issue?
On Sun, Apr 11, 2021 at 12:25 AM Lok P <loknath.73@xxxxxxxxx> wrote:
Below is the bug that seems to be exactly what we are suffering from
as it shows how "check point not completing" and "reliable message" are
associated with this BCT issue. In past we had suffered mainly from lot
of
"BCT Buffer space waits" and then we found BCT file size was 22GB and as
suggested by Oracle, we endup increasing the size of
_bct_public_dba_buffer_size from ~128MB to ~256MB and re-initialized the
BCT i.e. disbaled+enabled and that fixed the issue then. The size of the
BCT file after reinitialization was down to ~9 GB.
But it seems now within ~2months the BCT file has again grown back to
same ~22GB and this time we are seeing combination of waits like
"reliable
message" +"BCT buffer space waits", even with that
_bct_public_dba_buffer_size being set as ~256MB, so it seems this
doubling
of buffer size has not helped then rather re-initialization of BCT
helped.
So this time we are going to do the same reinitializing of BCT to have a
temporary fix.But now wondering what will be a permanent fix for this
issue?
High Waits On 'block change tracking buffer space' - Checkpoint
Contention With BLOCK CHANGE TRACKING or RMAN Incremental Backup (Doc ID
2094946.1)
On Sat, Apr 10, 2021 at 12:53 AM Lok P <loknath.73@xxxxxxxxx> wrote:
We have verified many occurrences and it looks like each time the
"reliable message" is followed by many "block change tracking buffer
space" wait events. We are thinking of initiating BCT by disabling +
enabling the BCT. but still unable to relate how come the BCT Buffer
space
can cause "reliable message" wait events making the database freeze?
