A Arpaci Dusseau CS739 Distributed Systems Department of Computer Science University of Wisconsin Madison Taming aggressive replication in the Pangaea wide area file system OSDI 02 1 Introduction What environment is Pangaea targeting What workload Is this well motivated What consistency model do they provide Why do they think this model is okay Why don t they want to provide strong data consistency 2 Pangaea A structural overview Where is your data stored in Pangaea Do you think this is a good idea What are the two purposes of replication How do Gold and Bronze replicas in Pangaea fulfill these different purposes What meta data must be tracked for each type of replica Why don t you want to make all copies Gold 3 Replica set management How does a user create a file Example User on server S creates file F in directory D How does a user on node S access an existing file F After S has a replica what must it do 4 Propagating updates Pangaea chooses to update replicas through flooding i e pushing the new contents to every neighbor Since the graph of replicas is strongly connected all replicas will eventually receive the updates What optimizations do they implement beyond flooding How does Pangaea determine whether or not it is correct to update a replica Updating directory entries is particularly interesting For example if simultaneously execute mv foo alice foo and mv foo bob foo hard to detect conflict since alice and bob might be stored on different nodes How does Pangaea solve this problem 1 If a user is worried about eventual consistency what can they do 5 Evaluation How much of the system is implemented What do they show in Table 2 and Figure 7 What is the point of Figure 8 What is the point of Figure 9 6 Class Discussion Questions What do you think is this major weakness of this work Given its goals and assumptions do you think Pangaea is designed well Why There is no central authority for determining who owns a file or who should host gold replicas What complications does this raise in Pangaea What do you think are the pros and cons of the allocation policy used in Pangaea Could Pangaea be modified to perform replication at the block level instead of the whole file level What complications would this raise Are there any disadvantages to hosting a bronze replica Why do you sometimes want to remove a bronze replica 2
View Full Document
Unlocking...