Fragmentation can cause serious loss of memory in systems that are using dynamic memory management. Any useful memory management system must therefore provide means to limit fragmentation. Today's garbage collector implementations often do this by moving objects in a way that free memory is non-fragmented. This paper presents a new object model that is based on fixed size blocks. The model eliminates external fragmentation without the need to move objects. A Java virtual machine and a static Java bytecode compiler that use this object model have been implemented and analysed using the SPECjvm98 benchmark suite. This Java implementation allows for deterministic memory management as needed in real-time systems that is difficult to achieve with moving collectors and unparalleled by current implementations.
Abstract. Root scanning is the task of identifying references to heap objects that are stored outside of the heap itself, in global and local variables and on the execution stack. Root scanning is particularly difficult within an incremental garbage collector that needs to be deterministic or give hard realtime guarantees. Here, a method that allows exact root scanning is presented. The basic idea is to ensure that copies of all root references exist on the heap whenever the garbage collector might become active. This approach reduces the root scanning phase of the garbage collection cycle to an efficient constant-time operation. A Java virtual machine and a static Java bytecode compiler that use this technique have been implemented and analysed using the SPECjvm98 benchmark suite. This Java implementation allows for deterministic memory management as needed in real-time systems that is difficult to achieve with traditional methods to perform root scanning.
More and more, parallel multicore systems will be used even in low-end devices such as embedded controllers that require realtime guarantees. When garbage collection is used in these systems, parallel or concurrent garbage collection brings important performance advantages. In the context of realtime systems, it has to be shown that a parallel garbage collector implementation not only performs well in most cases, but guarantees on its performance in the worst case are required.This paper analyses the difficulties a parallel mark-and-sweep garbage collector faces during a parallel mark phase. The performance of such a garbage collector degrades when only some of the available processors can perform scanning work in the mark phase. Whenever the grey set contains fewer elements than the number of available processors, some processors may be stalled waiting for new objects to be added to the grey set. This paper gives an upper bound for the number of stalls that may occur as a function of simple properties of the memory graph.This upper bound is then determined for the Java applications that are part of the SPECjvm98 benchmark suite and the theoretical worst-case scalability of a parallel mark phase is analysed. The presented approach is then applied to a Java virtual machine that has uniform mark steps, which at first results in poor worst-case scalability. A small change in the implementation is then proposed and analysed to achieve good scalability even in the worst case.
This paper provides an overview of the realtime garbage collector used by the RTSJ Java Virtual Machine JamaicaVM. A particular emphasis will be made on the improvements made in with release 3.0 JamaicaVM.The JamaicaVM garbage collector is incremental to an extreme extent: single incremental steps of the garbage collector correspond to scanning only 32 bytes of memory and have a worst-case execution time in the order of one µsec. The JamaicaVM garbage collector uses automatic pacing, making the system easier to configure than a garbage collector using explicit pacing that requires information on the application's allocation rate.The recent improvements of the garbage collector that will be present in this paper include support for automatic heap expansion; reduction of the memory overhead for garbage collector internal structures; and significant performance optimisation such as a faster write-barrier and a sweep phase that does not need to touch the objects and therefore reduces the number of cache misses caused during sweep.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.