High Performance MySQL cluster success story using Dolphinics SCI cards

[Roelof Kleinsmiede]

Hi Guys,

After having discovered clustered MySQL and Dolphinics SCI cluster interconnect cards through this Forum two years ago I thought it might be nice to contribute back to the community a success story that you all directly or indirectly helped to realise.

Just for the record, I work as a Lead Technical Architect for Capgemini in the UK and have no direct affiliation with either MySQL, Dolphin or IBM. My only motivation for writing this thread is to share a high performance design that works and is in Live use today!

Last year we went looking for a high performance database solution to act as a 'caching' front end to and existing national UK system. Our requirements were very stringent indeed: Design and build a solution to run on sub £100k hardware that could service up to 1,000 medium complex multi-read/write application transactions per second at an average response time of less than 0.1 seconds per application transaction and was capable of scaling to over 2,000. It also had to be resilient across multiple servers.

With your help the solution we ultimately chose comprises a pair of IBM p630 eServers with 16Gb RAM, 4 x 1.45Ghz Power4 CPU's running 64bit SuSE Enterprise 9 Linux and an NDB clustered MySQL v5. This was configured with 4 x 2Gb nodes per server with the cluster traffic configured to go across a pair of Dolphinics SCI high speed cluster interconnect cards. Quadruple LAN failure was provided by another pair of gigabit NIC's to back up the SCI cards.

Dolphin agreed to port their x86 SuperSockets drivers to PPC64 for us and we have been succesfully using these ever since.

The production application driven solution has been performance tested in a Live environment at 854 application transactions per second (= multiple read/write database records) via a Java/JDBC application and responds at 0.03 seconds on average per application transaction. Using an application test harness (i.e. not application driven) this solution was capable of 1,254 transactions per second at an average of 0.02 seconds response time. Our performance testing proved that in comparison to an active/active gigabit LAN cluster interconnect that the Dolphin SCI solution was responsible for over 35% of this performance vs. gigabit.

We are still constrained to using multiple nodes per server because in order to drop down to a more efficient single 16Gb node per server we need a MySQL supported 64bit binary distribution for PPC64 and to upgrade our SuSE Linux to Service Pack 3 to get >2Gb RAM per process. We have however tested our own 64bit MySQL PPC64 compilation and it works very well with 1 x 16Gb node per server. The point in mentioning it is that we estimate there is another 40% performance to be gained by dropping the number of nodes down from 4 to 1.

Hope this may be of some use to someone...