Software method level speculation for java

Speculative multithreading (SpMT), also known as thread level speculation (TLS), is a dynamic parallelization technique that relies on out-of-order execution, dependence buffering, and misspeculation rollback to achieve speedup of sequential programs on multiprocessors. A large number of hardware studies have shown good results for irregular programs, as have a smaller number of software studies in the context of loop level speculation for unmanaged languages. In this thesis we explore software method level speculation for Java. A software environment means that speculation will run on existing multiprocessors, at the cost of extra overhead. Method level speculation (MLS) is a kind of SpMT / TLS that creates threads on method invocation, executing the continuation speculatively. Although MLS can subsume loop level speculation, it is still a relatively unexplored paradigm. The Java programming language and virtual machine environment are rich and complex, posing many implementation challenges, but also supporting a compelling variety of object-oriented programs. We first describe the design and implementation of a prototype system in a Java bytecode interpreter. This includes support for various MLS components, such as return value prediction and dependence buffering, as well as various interactions with features of the Java virtual machine, for example bytecode interpretation, exception handling, and the Java memory model. Experimentally we found that although high thread overheads preclude speedup, we could extract significant parallelism if overheads were excluded. Furthermore, profiling revealed three key areas for optimization. The first key area for optimization was the return value prediction system. In our initial model, a variety of predictors were all executing naively on every method invocation, in order that a hybrid predictor might select the best performing ones. We developed an adaptive system wherein hybrid predictors dynamically specialize on a per-callsite basis, thus dramatically reducing speed and memory costs whilst maintaining high accuracy. The second area for optimization was the nesting model. Our initial system only allowed for out-of-order nesting, wherein a single parent thread creates multiple child threads. Enabling support for in-order nesting exposes significantly more parallelization opportunities, because now speculative child threads can create their own children that are even more speculative. This required developing a memory manager for child threads based on recycling aggregate data structures. We present an operational semantics for our nesting model derived from our implementation. Finally, we use this semantics to address the third area for optimization, namely a need for better fork heuristics. Initial heuristics based on online profiling made it difficult to identify the best places to create threads due to complex feedback interactions between speculation decisions at independent speculation points. This problem grew exponentially worse with the support for in-order nesting. Instead, we chose to clarify the effect of program structure on runtime parallelism. We did this by systematically exploring the interaction between speculation and a variety of coding idioms. The patterns we identify are intended to guide both manual parallelization and static compilation efforts.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.