>>>>> "RR" == Raymond Raymond <[EMAIL PROTECTED]> writes:
RR> Oystein wrote:
>>
>> I would like to suggest the following:
>> - 1. The user may be able to configure a certain recovery time
>> that Derby should try to satisfy. (An appropriate default
>> must be determined).
>> - 2. During initilization of Derby, we run some measurement that
>> determines the performance of the system and maps the
>> recovery time into some X megabytes of log.)
>> - 3. A checkpoint is made by default every X megabytes of log.
>> - 4. One tries to dynamically adjust the write rate of the
>> checkpoint so that the writing takes an entire checkpoint
>> interval. (E.g., write Y pages, then pause for some time).
>> - 5. If data reads or a log writes (if log in default location)
>> start to have long response times, one can increase the
>> checkpoint interval. The user should be able to turn this
>> feature off in case longer recovery times are no acceptable.
>>
>> Hope this rambling has some value,
>>
>> --
>> Øystein
>>
RR> Thanks for Oystein's comment. I agree with your comment
RR> and I have any other thought about it. In order to be easier
RR> to explain,I added the sequence number to your comment.
RR> In step 3 and 4 I have another idea. Generally, we do checkpointing
RR> from the earliest useful log record which is determined by the
RR> repPoint and the undoLWM, whichever is earlier, to the current
RR> log instant (redoLWM) and then update the derby control
RR> file(ref. http://db.apache.org/derby/papers/recovery.html).
I am not sure I understand what you mean by "do checkpointing". Are
you talking about writing the checkpoint log record to the log?
RR> I agree with
RR> you to spread the writes out over the checkpoint interval, but the
RR> trade-off is that we have to do recovery from the penultimate
RR> checkpoint(Am I right here?^_^). If the log is long, it will take us
RR> a long time in recovery.
>From the perspective of recovery, it will still be the checkpoint
reflected in the log control file. It is true that a new checkpoint
had probably been started when the crash occurred, but that may happen
today also. It is less likely, but the principles are the same.
I agree that it will be more log to redo during recovery. The
advantage with my proposal is that the recovery time will be more
deterministic since it will less dependent on how long time it takes
to clean the page cache. The average log size for recovery will
always be 1.5 checkpoint interval with my proposal. The maximum log
size will be 2 checkpoint interval, and this is also true for the
current solution. If the goal is to guarantee a maximum recovery
time, I think my proposal is better. It is no point in reducing
performance in order to be able to do recovery in 30 seconds, if the
user is willing to accept recovery times of 2 minutes.
RR> How about we update the derby control
RR> file periodically instead updating the control file when the whole
RR> checkpoint is done? (E.g. write several pages, if we detect that the
RR> system is busy, then we update the derby control file and pause for
RR> some time or we update the control file once every several
RR> minutes)
I guess that is possible, but in that case, you will need to have some
way of determining the redoLWM for the checkpoint. It will no longer
be the current log instant when the checkpoint starts. I guess you
can do this by either scanning the entire page cache or by keeping the
pages sorted by age.
RR> That seems we do a part of checkpoint at a time if the system
RR> become busy. In this way, if the system crushes, the last checkpoint
RR> mark (the log address up to where the last checkpoint did)will be
RR> closer to the tail of the log than if we update the control file when
RR> the whole checkpoint is done. Maybe we can call it Incremental
RR> Checkpointing.
Unless each checkpoint cleans the entire cache, the redoLWM may be
much older than the the last checkpoint mark. Hence, updating the
control file more often, does not reduce recovery times by itself.
However, making sure that the oldest dirty pages are written a
checkpoint, should advance redoLWM and reduce recovery times.
--
Øystein
