Memory and Object Management in Ramcloud a Dissertation Submitted to the Department of Computer Science and the Committee on Graduate Studies of Stanford University in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy

Traditional memory allocation mechanisms are not suitable for new DRAM-based storage systems because they use memory inefficiently, particularly under changing access patterns. In theory, copying garbage collectors can deal with workload changes by defragmenting memory, but in practice the general-purpose collectors used in programming languages perform poorly at high memory utilisations. The result is that DRAM-based storage systems must choose between making efficient use of expensive memory and attaining high write performance. This dissertation presents an alternative, describing how a log-structured approach to memory management – called log-structured memory – allows in-memory storage systems to achieve 80-90% memory utilisation while still offering high performance. One of the key observations of this approach is that memory can be managed much more efficiently than traditional garbage collectors by exploiting the restricted use of pointers in storage systems. An additional benefit of this technique is that the same mechanisms can be used to manage backup copies of data on disk for durability, allowing the system to survive crashes and restarts. This dissertation describes and evaluates log-structured memory in the context of RAMCloud, a distributed storage system we built to provide low-latency access (5μs) to very large data sets (petabytes or more) in a single datacentre. RAMCloud implements a unified log-structured mechanism both for active information in memory and backup data on disk. The RAMCloud implementation of log-structured memory uses a novel two-level cleaning policy, which reduces disk and network bandwidth overheads by up to 76x and improves write throughput up to 5.5x at high memory utilisations. The cleaner runs concurrently with normal operations, employs multiple threads to hide most of the cost of cleaning, and unlike traditional garbage collectors is completely pauseless.