pbxt-discuss team mailing list archive

[Paul McCullagh <paul.mccullagh@xxxxxxxxx>]

Hi Erkan,

As you have discovered already. The best way to see what PBXT is doing  
is to use xtstat.

This has been improved over the various versions. 1.5 gives the most  
information.

On Nov 27, 2010, at 2:54 PM, erkan yanar wrote:

> Hi Paul,
> (sorry for hijacking the thread)
>
> On Sat, Oct 30, 2010 at 06:19:59PM +0200, Paul McCullagh wrote:
> > Also, writing an index is "all-or-nothing" so since about 20 GB of
> > index data is dirty it probably all has to be written.
> >
> > Of course, as soon as one thread runs into this problem so do they  
> > all.
> >
> > Normally this situation should be avoided...
>
> So is there a tip how to avoid these problem (with 1.0)?
> Being still a noob about PBXT and I can't think about something else
> than increasing pbxt_index_cache_size. Doing some testing comparing
> massive parallel INSERTS (concurrency 8), the cache get filled faster
> than maybe any single thread (sweeper?) could clean it.

UNfortunately, in 1.0, the only way to improve this is to increase the  
index cache size.
> I missed the function about ilog-1.xt. From the xtstat output it is  
> correlated to
> the indexfiles. But I missed the way it works (is configured) exactly.

I did not describe this in my talk. The ilog are used to do a  
consistent flush on the index files.

All changes to an index file are first written to an ilog. Then the  
changes in the ilog are applied to the index file.

On recovery, the system checks that all changes to the index file are  
either applied or none at all.

This ensures the physical consistence of index files. But not the  
logical consistency, which is recovered by normal recovery and action  
of the sweeper.

> Is there a way to know if the Sweeper,Writer, etc. Threads are  
> running?
> As in my observations PBXT tends to stress the disks more than InnoDB.
> So it would be nice to know what is doing the workload at that time.

You should be able to see this in xsstat.

> On some tests all threads get stucked, which could be because of the
> index-thing (mentioned in the quoted text) or something else.
>
> This leads me to a feature-assumption:)
> I would like to have something like PBXT_TRX, PBXT_LOCKS etc. where we
> could also see the Sweeper,Writer etc. threads.
>
> Please excuse my silly questions, as I still claim the freedom of a  
> noob:)

All questions are welcome!

Best regards,

Paul



> --
> über den grenzen muß die freiheit wohl wolkenlos sein
>
> _______________________________________________
> Mailing list: https://launchpad.net/~pbxt-discuss
> Post to     : pbxt-discuss@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~pbxt-discuss
> More help   : https://help.launchpad.net/ListHelp