How can I increase one thread's throughput

[Tom Shen]

We are do PoC on MySQL Cluster and try to find how can we migrate to it for our application. Our old data persist tier is very fast (one thread can do 80K database operations per second) but it's not friendly for horizontal scalability.

I change ndbapi_simple.cpp (only do one record insert in one transaction) do the similar test on i7-2600 with NoOfReplica=1 and find the following (the number is similar when I choose either ndbmtd or ndbd
Thread Number, Throughput per Thread
1, 19K
4, 11K
10, 7K
20, 5K

When I change to two machine with 2xE5-2670 (8 cores per socket) and 64G memory, this time NoOfReplica=2 (I use ndbmtd and MaxNoOfExecutionThreads=8 for this)
Thread Number, Throughput per Thread
1, 3.3K
4, 3.9K
8, 3.6K
16, 3.1K
32, 2.7K
64, 2.3K
128, 1.8K
256, 1.4K

I'm not surprised for the throughput drops so much for NoOfReplica=2. I found a lot of thread are required for fill one data group. Which will make he business layer must have a lot of thread than before.

I use localhost as connection string.

My Question is how can I increase one thread's throughput for one insert per transaction.