[brianatdts]

the problem with scaling out to multi cores with a focus on ram is that in larger datasets you end up trading disk latency for network and protocol latency. I a not sure that is a great trade even if we are talking about fiber channel as a medium.

[dennis82]

I have to disagree; disk is ancient - it's mechanical egads! - while 10GigE is pretty commonplace now and infiniband and fiber channel are even faster.

back from my CS 101 takeaways: there are only 3 bottlenecks in a computer system: CPU, network, and IO.

looks like MemSQL is fixing the CPU and IO bottlenecks, but physics is physics so network is pure hardware solution haha

[simcop2387]

The problem is that you can end up with larger latency over the network because it still takes a fixed amount of time for nodes to communicate. Even with a 1TB/s link between nodes you can still have a good 30ms between them all adding even more latency. That can be mitigated somewhat by a good protocol that can manage that latency properly (e.g. not blocking while waiting on ACKs and such), it can still end up with far more latency than a few large disks would be (even better now with SSDs). That said I do imagine that some datasets will benefit from this kind of topology (I can imagine that geospatial stuff will do well with that, since you can locate physically close things on a single machine and reduce the amount of talking needed).

[TylerE]

30ms? In anything resembling a modern datacenter? 0.3-0.5ms is more typical these days.

[brianatdts]

He was joking.