MySQL Forums
Forum List  »  MySQL & Storage

Re: MySQL on ZFS - transactional filesystems = fragmentation
Posted by: Nils Goroll
Date: October 30, 2007 08:10AM

First of all, thanks to John for this *excellent* article. I am happy to see a benchmark showing that ZFS can now be considered even for use with InnoDB. This all looks very promising, but despite the fact that I believe that ZFS is an ingenious innovation by Sun (and despite the fact that I am a strong advocate for Sun in general and Solaris in particular), I fear that the setup could be favoring ZFS:

- It would be interesting to see results for a larger working set / available memory ratio - iow what if the working set is not just 3 GB for 2 GB ram, but maybe 10 fold larger.

- It would also be interesting to see how the read/write ratio of the benchmark affects the results.

I would also like to add one aspect which I believe is significant for production Systems: John has pointed out that transactional file systems / storage can make obsolete the doublewrite buffer. But nothing comes for free, as John has also mentioned, transactional file systems, "never" overwrite data but rather write new blocks "somewhere else" instead (I like the notion of "versioned blocks").

This means that locality of data within a datafile is not mapped to physical storage at all. While with traditional file systems or raw devices, a sequential read on a datafile will in most cases end up in a sequential read on physical storage, on transactional file systems it will most likely end up in a random read. And, even worse, this additional layer of fragmentation degrades over time with the number of writes committed to the storage, so one would also need to benchmark over time to get proper results.

To make a long story short, with transactional file systems we don't just have to worry about fragmentation in the datafile, but also about additional fragmentation in the file system / storage - or we have to size systems assuming that any I/O will be random I/O.

Please note that these are theoretical thoughts which I have not have taken the time to verify or falsify in practice. If you know any benchmarks on this effect, please drop me a note.

Finally, thanks again for the great article,

Nils



Edited 1 time(s). Last edit at 10/30/2007 08:11AM by Nils Goroll.

Options: ReplyQuote


Subject
Written By
Posted
October 25, 2007 07:33PM
October 25, 2007 09:09PM
Re: MySQL on ZFS - transactional filesystems = fragmentation
October 30, 2007 08:10AM


Sorry, you can't reply to this topic. It has been closed.

Content reproduced on this site is the property of the respective copyright holders. It is not reviewed in advance by Oracle and does not necessarily represent the opinion of Oracle or any other party.