Re: Test cluster slower than single machine
This benchmarks simulates students taking an on-line test, and the predominant query in this benchmark is a one-row INSERT into the bench_response table. This table has an auto_increment primary key, plus an ordered index on bench_instance (which identifies something like a student session).
I suggest starting the MySQL nodes with
ndb_autoincrement_prefetch_sz = 250
ndb_use_exact_count = 0
Brad, you should be able to pretty clearly show that every time you add a MySQL node, the performance goes up. I would definitely expect to see this happen going from 1 mysql server to 2, and from 2 to 3, and probably all the way up to 5.
Then, once it stops increasing, if you add NDB nodes (i.e. more fragments), they should support more MySQL nodes so you can keep going up until eventually the network becomes the bottleneck.
On an SMP machine you can pretty much have NDB node per CPU -- so it sounds like you could put several fragments on that SGI box.