Moore's law, has spoiled us, instead of inward looking we say Heil Moore's law.
I remmber, the Hadoop wave was buy less expensive hardware and today whats a comodity server Here are the recommended specifications for DataNode/TaskTrackers in a balanced Hadoop cluster:
- 12-24 1-4TB hard disks in a JBOD (Just a Bunch Of Disks) configuration
- 2 quad-/hex-/octo-core CPUs, running at least 2-2.5GHz
- 64-512GB of RAM
- Bonded Gigabit Ethernet or 10Gigabit Ethernet (the more storage density, the higher the network throughput needed) Seriously ????
If thats not enough Our findings suggest SSD has higher performance compared to HDD-11
Does it stops there, no; there is something we could do networking as these are super chatty, hence lets bring in InfiniBand
World is not enough (007), you know what CPU is not enough bring in GPU
Now this is our simple and humble less expensive comodity server ;)
I think its time, we either look for something else or crush the big-data stack; there are new buds in form or kudu... but I feel crushing the stack is the next logical step. What if we make NN,RM,DN run as single process inside a unikernel. In an year or so the unikernel will be a viable tech and if we are ready by then the current performance number will be microscopic.
This project is a humble start towards that direction; Where no man has gone before