Continuent Community

Dear all,

I'm now testing Tungsten Enterprise to see if we as a company can use Tungsten to solve our future database scale out problem.

There are generally three scenario's you can have with a database:
1) read heavy
2)read and write equally heavy
3) write heavy

I'm now benchmarking bare metal and a Linux KVM instance with sysbench http://sysbench.sourceforge.net/
The benchmark tool uses a Mysql C client api to connect to a database and fire off sql commands.

can i use this tool also against Tungsten cluster? (as far as i know only JDBC...)

I want to compare bare metal vs KVM vs a tungsten cluster.

the second question is, how well does tungsten handles read and write equally heavy and write heavy databases? is there a general advice about this?

Thanks in advance,

Damir

Hi Damir,

You can definitely use sysbench against Tungsten. Just point it straight at the master database. One thing I should point out up front is that Tungsten is slower than MySQL replication because we re-log transactions in InnoDB to capture them safely. This ensures we have better recovery behavior than MySQL.

We have optimizations in Tungsten 1.3 that will improve performance and will be introducing a new log format as well as parallel replication that will make Tungsten quite a bit faster than MySQL.

Meanwhile, have fun with testing. Also, look at the performance settings in the replicator.properties--these can speed up replication quite a bit.

Cheers, Robert

I shall await with eager anticipation to watch your times tumble
Good luck with the new pedals I am sure that they will make racing more pleasurable
________________________________________________________________________________ __________

Noise Cancelling Headphones| SEO Agency

Robert Hodges wrote:

QUOTE:

Hi Damir,

You can definitely use sysbench against Tungsten. Just point it straight at the master database.
...
Cheers, Robert

Since Tungsten working on JDBC level, when using MySQL C client API to write to master database, will transaction be fully replicated to slave database? If so, how does it work?
Will "zero data loss" still be guaranteed when there's no DRBD like HA support? If so, how will it ensure data guarded?

Thanks,
Robin

Hi Robin!

Tungsten ensures zero data loss in the case of a planned failover (done through the cctrl 'switch' command) by shutting down connectivity and pushing a final flush transaction through the system to ensure the slave is up to date.

In the case of unplanned failover, it is more complicated, as you may have data that never gets off the master to the slave that is promoted on failure. In this case the data are not exactly 'lost' but more like 'a little hard to find'. You can find out pretty quickly what is left on the master using the Tungsten thl utility, which can dump data starting at a specific transaction ID. You then either turn this into a SQL script or manually reconcile through the application if you need the transactions.

We don't currently support synchronous replication, as this has pretty steep costs in terms of application availability, and requires support from the MySQL server that will not be available until MySQL 5.5.

Cheers, Robert