9.21.1. Checkpoint & Page Remapping
In concept, making a checkpoint consists of writing the last committed state of a page on the read-only pre-checkpoint state. This is in case the page existed before the previous checkpoint. If the page was created after the last checkpoint, making the checkpoint consists of just writing it to disk. Now if the number of pages modified between checkpoints largely exceeds the amount of RAM, the checkpoint will be a disk-to-disk copy process which takes a time proportional to the number of modified pages. This can take a long time. Virtuoso offers a mechanism called checkpoint remap. This allows making a committed state persistent, i.e. safe, without copying all the delta collected since the last checkpoint over the pre-checkpoint state.
The checkpoint remap mechanism means that a page, say number 12 get remapped to 15 when updated. Suppose now that the page were written to disk. Now we have the pre-checkpoint page on 12 and the committed post checkpoint state on 15. If a checkpoint now takes place, instead of copying 15 over 12, which may involve reading the page, we just write that 12 is actually on 15. This speeds up the checkpoint but causes one page to take the space of 2. Now suppose 12 were modified again, now we would see that the pre checkpoint page is 15 and that the original 12 is free. The page now gets mapped to 12. The next checkpoint now will mark that 12 is 12, which was the original condition and 15 is free.
The mechanism is actually more complex but this is the basic idea. By allowing a page to occupy two pages in some situations we dramatically cut down on atomic checkpoint time and improve availability.
Now we may argue that this destroys locality since pages that were created contiguous are no longer that after an update. This is why there is also a mechanism called unremap. and there is an upper limit to checkpoint remaps. If every page were updated between checkpoints, they would alternately be in their original place or in a remote place from which the next update-checkpoint combination would return them to the original. Now this is not always the case. Also there is not always enough space to keep up to three copies of each page. Therefore there is an unremap mechanism and an upper limit to checkpoint remaps.
parameter in the virtuoso.ini file controls how
many pages may be stored on a page other than their logical page.
in the virtuoso.ini file controls how
many previously remapped pages are unremapped each time a
checkpoint is made.
Having an unlimited MaxCheckpointRemap and zero UnremapQuota will cause a checkpoint to consist exclusively of a flush of dirty buffers. The downside is that a page can take the space of two for an unlimited period of time and that original locality is not maintained. Setting the maximum checkpoint remap to zero in such a situation causes the next checkpoint to unremap all checkpoint remapped pages, restoring physical sequence and freeing disk space.
The TPC C Benchmark chapter for examples of checkpoint remapping in use on test environments.