Does this number factor in the IO operations? Or is it just the number of transactions the architecture can match reading and writing only to/from memory?
Which hardware did they use during the test? It probably doesn't matter but which OS did they use?
With LMAX, IO is usually out of the transaction loop and done somewhere else.
Until somebody can answer who has actually run that benchmark, I can only assume they had some kind of counter and some kind of time measurement. But I don't think this was on a multi-socket configuration.
Other than taking this 6M TPS hearsay from 2011 or 2010 for granted, you should run your own tests with Disruptor. The benchmark they ran may or may not reflect what you need to accomplish, and unfortunately, these latency/throughput numbers will vary a lot for the slightest change in circumstances. Knowing the OS or the hardware may statisfy your curiosity, but will not make the numbers any more meaningful than your own evaluation.