Re: poor performance on load data to NDB

[Kevin Sigrist]

Hi,

You've got to remember that any operation within the Cluster normally have numerous round trips across the network as messages are passed between the data nodes and MySQL. Direct comparissons with myisam or single NDB node response times aren't applicable because of the unique architecture of cluster.

Significant amounts of traffic are transferred between the servers in the cluster, so any latency between nodes is a bad thing for your response times.
The number of messages will increase as you introduce data nodes and also factor in NoOfReplicas. All inserts from your load will have to be committed across NDB via two phase commit.

What sort of interconnects are you running?, in terms of repsonse time and throughput one of the most important physical aspects is going to be your network interconnect.

I assume you are connected via TCP/IP. So what speed NICs are you running? Gigabit I hope. Ideally all your nodes are also connected via a dedicated switch (Gigabit as well) devoted just to your cluster traffic. (this is recommended for performance as well as security). Putting the data nodes on your existing network is going to cause trouble.

As an aside, The SCI cluster interconnects are the current nirvana in terms of response. I've seen MySQL AB quoting 50% reduction in response time for network bound queries.

Finally you might want to have a look at the variable ndb_force_send. Turning it off may improve response time. Turning it on will improve throughput.

Regards
Kevin