CS 553 Internet Services Fall 2015 Questions on Service Architecture readings, part 1 A. Paper: Brewer, et. al., Lessons from Giant-Scale Services. http://www.cs.berkeley.edu/~brewer/Giant.pdf 1. Is author's claim that services have a high asymmetry of read vs. writes still true today? Why or why not? 2. What are additional features/properties that Internet Services must support in addition to high availability? 3. Why kinds of trade-offs do the harvest and yield metrics allow services to describe compared to traditional reliability metrics? Give an example of one-such engineering design trade-off that is easily articulated using harvest or yield. 4. Suppose an Internet Service's output was specific to each user, e.g. Facebook, or a banking site. Would the DQ principle still apply? Give an example showing it applies, or an example where it does not. 5. Describe one constraint of the incremental on-line evolution strategies, i.e., rolling upgrades and the big flip, that the fast reboot approach does not have. That is, under what conditions are the incremental strategies more challenging? B. Paper: Armbrust, et. al., A view of Cloud Computing http://www.cs.berkeley.edu/~rxin/db-papers/cloudcomputing.pdf 1. What is the authors' definition of cloud computing? 2. Summarize the authors' use-cases for cloud computing. 3. Explain under what circumstances a cloud computing system could be preferable to owning your own servers even when buying and depreciating your servers is cheaper. 4. Suppose the peak server load exceeds the average workload by a factor of 6, and the peak load occurs 10% of the time, what would be the average server utilization? You can express load in a percentage of web requests/second. 5. Of the authors' list of obstacles for cloud computing, which do you think are the most important 3 to address, and why? C. Paper: Web Search for a Planet: The Google Cluster Architecture http://static.googleusercontent.com/media/research.google.com/en//archive/googlecluster-ieee.pdf 1. What is a key metric the authors' claim is very important, which was omitted in the Brewer paper? 2. Suppose we need to increase the number of replicas to decrease the latency of serving requests. What other service metrics should also improve by increasing the number of replicas? 3. List one advantage of a strategy that uses many smaller nodes/computers in a cluster over a fewer larger nodes/computers. Where might the size advantage of a smaller node break down? That is, when would a node become "too small"? 4. Compare the power density (energy per time expended by volume) in an Internet Service cluster to the core of the sun. Which is greater and by how much? Justify your answer.