Adaptive optimization in the Jalapeno JVM

Abstract

Future high-performance virtual machines will improve performance through sophisticated online feedback-directed optimizations. This paper presents the architecture of the Jalapeno Adaptive Optimization System, a system to support leading-edge virtual machine technology and enable ongoing research on online feedback-directed optimizations. We describe the extensible system architecture, based on a federation of threads with asynchronous communication. We present an implementation of the general architecture that supports adaptive multi-level optimization based purely on statistical sampling. We empirically demonstrate that this profiling technique has low overhead and can improve startup and steady-state performance, even without the presence of online feedback-directed optimizations. The paper also describes and evaluates an online feedback-directed inlining optimization based on statistical edge sampling. The system is written completely in Java, applying the described techniques not only to application code and standard libraries, but also to the virtual machine itself.

Characterizing and predicting program behavior and its variability

“…We obtain a better measure of variability by considering the average absolute distance between adjacent points (mean deltas) in the time series. Figure 3 shows the mean deltas in IPC values between adjacent time intervals for each of the programs and for both the Power3 and the Power4 1 . While we present the deltas only for IPC, similar behavior variability is seen across all metrics.…”

“…Another example is memory management in Java Virtual Machines (JVM), where the garbage collector's activities are adapted to the application's behavior [7,23]. Just-In-Time compilers, such as the ones found in JVMs [1,2], and dynamic optimization frameworks [4,19] target adaptation of the executing code itself in order to exploit dynamic program characteristics.…”

Correcting the Dynamic Call Graph Using Control-Flow Constraints

“…Programs often decompose functionality into small methods, and virtual dispatch often obscures call targets at compile time. The dynamic call graph (DCG) records execution frequencies of call site-callee pairs, and dynamic optimizers use it to analyze and optimize whole-program behavior [2,3,4,5,11,21]. Prior approaches sample to collect the DCG, trading accuracy for low overhead.…”

“…Most dynamic optimizers collect accurate control-flow profiles such as basic block and edge profiles to make better optimization decisions [1,3,12,16,17]. We show how to combine these constraints to further improve the accuracy of the DCG.…”

“…We evaluate DCG correction in Jikes RVM [3] on the SPEC JVM98 and DaCapo [8] benchmarks. Our approach improves DCG accuracy over the default sampling configuration in Jikes RVM, as well as over the counter-based sampling (CBS) configuration recommended by Arnold and Grove [4].…”