CSE 124 Lecture Notes - Lecture 28: Delimiter, Vector Clock, Cloud Computing
9 Jun 2018
School
Department
Course
Professor
Power Proportional Humans
Web-Service Load Fluctuations
- Some components may try to “reduce” the power consumption (depending on the load)
Do Different Components Scale Similarly?
- in a perfect world, we want the amount of traffic is able to handle the loads
- reducing power usage in components is not particularly effective
- the amount of power that the processor uses for bigger workloads is not that much higher!
- it’s REALLY difficult to run disks and RAM with “low” power
- ideally, our distribution between compute load and power consumption would follow y = x!
- but realistically it is NOWHERE close
- it is actually better to run your systems at close to 100% utilization!
What About “Power Saving” Features on Modern Computers
- once you start “scaling” things out on many machines, the power saving is not as “impactful”
- it’s better to just keep your systems HIGHLY utilized!
Power Takeaway
- cloud computing has HUGE energy/cost implications
- efficiency via high utilization
- pack MORE workloads for highly utilized systems
Traditional DC Topology
Internet
---------------------------------
Core Layer-3 routers
Aggregation
Access Racks/Switches
Tree-based network topologies
- as servers got faster, it became harder to “build” these networks!
- the bi-section bandwidth at the top of the Tree would have to forward a 1 Pb of data PER
second!
Capacity Bottlenecks
- if ALL (n) servers are busy, each one only gets 1/n the bandwidth
- the idea of forwarding traffic up to the roots and back causes TOO many bottlenecks!
Folded-Clos multi-rooted Trees
- use MANY switches throughout the network
- allow for MANY different paths to send data
- each server can now (potentially) run at full utilization!
Document Summary
Some components may try to reduce the power consumption (depending on the load) In a perfect world, we want the amount of traffic is able to handle the loads. Reducing power usage in components is not particularly effective. The amount of power that the processor uses for bigger workloads is not that much higher ! It"s really difficult to run disks and ram with low power. Ideally, our distribution between compute load and power consumption would follow y = x ! It is actually better to run your systems at close to 100% utilization! What about power saving features on modern computers. Once you start scaling things out on many machines, the power saving is not as impactful . It"s better to just keep your systems highly utilized ! Pack more workloads for highly utilized systems. As servers got faster, it became harder to build these networks!
