Certainly, default (small) temp buffer size plays roles. But it this IPC host this difference is not so important. Result with local temp tables and temp_buffers = 1GB: 859k TPS.
It is little bit unexpected result.I understand so it partially it is generic problem access to smaller dedicated caches versus access to bigger shared cache.
But it is hard to imagine so access to local cache is 10% slower than access to shared cache. Maybe there is some bottle neck - maybe our implementation of local buffers are suboptimal.
Using local buffers for global temporary tables can be interesting from another reason - it uses temporary files, and temporary files can be forwarded on ephemeral IO on Amazon cloud (with much better performance than persistent IO).
