Cloud computing

Data warehouse
Energy consumption
Electricity output 2X 2002~2007

Ware-house data center

efficiency / centralize / scaling 

K-means clustering

Method:

隨機放用k個centroids,每個centroid向最接近它的所有item的平均位置移動,移動後再重新計算,直到所有的centroid都沒在移動後就結束

K-means clustering

WSC (warehouse-scale computer)
Aggregation(規模經濟) - cost down


- Platform level software: Kernel
- Cluster-level infrastructure: Distribution File Systems, RPC layers.
- Application-level software: online service and offline computations(Google earth).


Datacenter vs Desktop
- Ample parallelism
data-request
- Workload churn
deploy more quickly
rapid product innovation
costumer focus
short software development cycle
- Platform homogeneity (same power supply, package)
- Fault-free operation


* Datacenter vs Desktop
decrease system complexity


Load-balancing -> cost down power
active -> DVFS
transparency -> load balancing




Application-level Software
reorganization: data-center > algorithm


Server power vs. utilization
power(f, u) = f + u(1-f)


saving time



switch on event
idle -> switch to another thread
CPU (DVFS) -> not easy to power consumption
DRAM -> important issue


Performance weight > Watt weight
Performance / Watt