DiskJournal performance

classic Classic list List threaded Threaded
4 messages Options
Reply | Threaded
Open this post in threaded view
|

DiskJournal performance

alaley
Hello,

are there any performance tests for DiskJournal exist?
I'm interesting on how much TPS can be achieved with it,
assuming almost zero delay commit time for certain datasource (faked).

Thanks,
Serge.

---------------------------------------------------------------------
To unsubscribe from this list, please visit:

    http://xircles.codehaus.org/manage_email


Reply | Threaded
Open this post in threaded view
|

Re: DiskJournal performance

Ludovic Orban
Administrator
Hi,

Well, there are no published performance tests nor benchmarks available for the DiskJournal. All I can tell you is that I've heavily tested it and compared it against comparable competing solutions and the performance was pretty much the same.

The real bottleneck is the disk syncing time: each time a transaction is committed the changes made to the log files must be flushed to disk which takes quite some time. Batching helps a lot with performance when you're multithreading but in the end, you disk speed is the true bottleneck.

To give you an idea, a cheap disk like the one you find in any random workstation can sustain something in the order of 50 disk writes per second, a fibre-channel SAN can do much more.


Is that performance a concern to you ?

Ludovic
Reply | Threaded
Open this post in threaded view
|

Re: DiskJournal performance

alaley
Ludovic Orban пишет:
>
> To give you an idea, a cheap disk like the one you find in any random
> workstation can sustain something in the order of 50 disk writes per second,
> a fibre-channel SAN can do much more.
> Is that performance a concern to you ?
Yes, it is. Thank you for providing some concrete numbers.
Hope proper RAID configuration with modern server HDDs
will do a good job too.

Do you think it have sense to perform regular backup
of BTM transaction logs during long-run application?
And is there any way to backup tx logs in atomic manner,
like "svnadmin dump" (to get a consistent snapshot of data
via non-blocking operation)?

Thanks,
Serge.

---------------------------------------------------------------------
To unsubscribe from this list, please visit:

    http://xircles.codehaus.org/manage_email


Reply | Threaded
Open this post in threaded view
|

Re: DiskJournal performance

Ludovic Orban
Administrator
Sure, RAID will definitely help the performance of disk syncs but please do not forget that unless you really require very high throughput and have extremely fast databases / JMS servers the transaction manager's performance isn't an issue most of the time.

Your backup question is a good one. There is no mechanism to make hot backups of transaction logs built in BTM so keeping the logs on a safe disk (like a RAID array) definitely is a wise idea.

Finally, there is a way to safely run XA transactions without the need of a disk journal: Logless Transactions which fixes both problems: a disk store isn't needed anymore so the disk's performance isn't an issue anymore and there is no need to backup the disk logs. Unfortunately BTM does not implement it yet because I don't have the resources to invest in that development.

Ludovic