Availability analysis

This section studies the effects of pervasive replication, especially name-space containment, on the system's availability. A Pangaea server replicates not just replicas accessed directly by the users, but also all the intermediate directories needed to look up those replicas. Thus, we expect Pangaea to disrupt users less than traditional approaches that replicate files (or directories) on a fixed number of nodes.

Figure 16: Availability analysis using a file-system trace; the users of a failed node move to a functioning node. The numbers in parentheses show the overall storage consumption, normalized to pang-1.

We perform trace-based analysis to verify this prediction. Two types of configurations are compared: Pangaea with one to three gold replicas per file, and a system that replicates the entire file system contents on one to four nodes. Our trace was collected on our departmental file server, and it contains 24 users and 116M total accesses to 566K files [31]. To simulate a wide-area workload from this single-node trace, we assume that each user is on a different node; thus, all the simulated configurations contain 24 nodes.

For each configuration, we start from an empty file system and feed the first half of the trace to warm the system up. We then artificially introduce remote node crashes or wide-area link failures. To simulate the former situation, we crash 1 to 7 random nodes and redirect accesses by the user on a failed node to another random node. To simulate link failures, in which one to four nodes are isolated from the rest, we crash 20 to 23 random nodes and throw away future activities by the users on the crashed nodes. We then run the second half of the trace and observe how many of the users' sessions⁷ can still complete successfully. We run simulation 2000 times for each configuration with different random seeds and average the results.

Figure 16 shows the results. For network partitioning, Pangaea wins by a huge margin; it shows near-100% availability thanks to pervasive replication, whereas the other configurations must rely on remote servers for much of the file operations. For node failures, the differences are smaller. However, we can still observe that for the same storage overhead, Pangaea offers better availability.