Michael Franz. Run-time code generation as a central system service. In Proceedings of the fifth Workshop on Hot Topics in Operating Systems (HotOS-VI), pages 112--117, May 1997. (p 54) 181  Home/Search   Document Not in Database   Summary   Related Articles   Check

....from developments in optimization technology. Similarly, if compilation is delayed until run time then value based or cross module specialization may be possible. Franz advocates such a system in which code from various sources is combined dynamically into a single quasi monolithic image [Franz97] However, one of the key tenets of the virtual machine proposed in Chapter 4 is that such concerns may be addressed by separating the underlying application from its specialized execution policies thereby retaining flexibility for both the vm implementor and the application programmer. Loss ....

Michael Franz. Run-time code generation as a central system service. In Proceedings of the fifth Workshop on Hot Topics in Operating Systems (HotOS-VI), pages 112--117, May 1997. (p 54) 181

....for use with files opened exclusively for reading [23, 8] Franz has also proposed making run time code generation a central operating system service. In his system, mostly compiled code is compiled just in time when an application is launched, then recompiled for optimization as the program runs [13]. Specialization is partial evaluation working in the other direction from extensibility. When a code path is specialized, it is no longer as general and flexible as before, but now has fewer tests and branches, and so executes faster. Because the partial evaluator is constantly refining the ....

M. Franz. Run-time code generation as a central system service. In Proceedings of the Sixth Workshop on Hot Topics in Operating Systems (HotOS), pages 112--117, 1997.

....(customised instructions and primitives) eases the implementation of a tightly integrated MOP (for introspection and intercession) and facilitates runtime behavioural veri cation. EOEciency can be assured by including the nal stage of compilation as an OS service (dynamic code generation) [Fra97]. Merging these two approaches (system language integration and virtual machine based execution) yields what we call a virtual execution environment (VEE) The rst such systems are starting to appear, in the form of JanOS for example [JANOS] The major remaining problem is that the system ....

M. Franz, Run-Time Code Generation as a Central System Service, Proc. HotOS-VI,May 1997. http://www.ics.uci.edu/~franz/publications/

....it is with a linear byte code sequence. 3. A Run Time Architecture Featuring Dynamic Re Optimization We are developing a run time architecture in which the capability of generating executable code from a portable intermediate representation is a central function of the operating system itself [Franz 1997b] It thereby becomes possible to perform advanced optimizations that transcend the boundaries between individual portable components, as well as the boundary between user level and system level code. Consider a scenario in which a user downloads several portable components from various Internet ....

M. Franz (1997); "Run-Time Code Generation as a Central System Service"; in The Sixth Workshop on Hot Topics in Operating Systems (HotOS-VI), IEEE Computer Society Press, ISBN 0-8186-7834-8, 112-117.

....small. Moreover, we use a smart branch and bound variant of the algorithm that is an order of magnitude more e#cient than a naive implementation. 5. IMPLEMENTATION CONTEXT We have implemented a version of the described algorithm and integrated it into our dynamic code generation infrastructure [Kistler 1997; Franz 1997; Kistler and Franz 1999] This infrastructure consists of a system that continually profiles all executing code, dynamically generates globally optimized versions of the same code in the background and then hot swaps the optimized code image in place of the previously executing one. The ....

....a series of static optimizations on critical procedures and instruments them with path profiling code. The performance of the resulting executable is presented in Figure 4 under the heading original layout with 2 Our particular implementation uses the Slim Binary representation [Franz and Kistler 1997], but our solution does not depend on this fact. Our system could be adapted to use programs represented as class files for the Java Virtual Machine, or even native code for some specific processor. This would merely have an e#ect on the pre processing e#ort required to extract information ....

T. Kistler and M. Franz Franz, M. 1997. Run-Time Code Generation as a Central System Service. In Proceedings of the Sixth Workshop on Hot Topics in Operating Systems (HotOS-VI), pp. 112--117. IEEE Computer Society Press.

....we use a smart branch and bound variant of the algorithm that is an order of magnitude more e#cient than a naive implementation. 5. IMPLEMENTATION CONTEXT We have implemented a version of the described algorithm and integrated it into our dynamic code generation infrastructure [Kistler 1997; Franz 1997; Kistler and Franz 1999] This infrastructure consists of a system that continually profiles all executing code, dynamically generates globally optimized versions of the same code in the background and then hot swaps the optimized code image in place of the previously executing one. The ....

....optimized for one particular client performs when it is used with another one. To this e#ect, we performed a series of 4 These break even results assume that re optimization occurs in the background on the same processor that is executing user programs. As one of the authors has argued elsewhere [Franz 1997], this scenario may actually turn out to be unnecessarily pessimistic. It may very well be cost e#ective to delegate the task of re optimization to an additional processor, so that the CPU running the user s application is not burdened by the re optimization task. The reasoning is as follows: ....

Franz, M. 1997. Run-Time Code Generation as a Central System Service. In Proceedings of the Sixth Workshop on Hot Topics in Operating Systems (HotOS-VI), pp. 112--117. IEEE Computer Society Press.

....TDB. Automated Data Member Layout of Heap Objects 9 is an order of magnitude more e#cient than a naive implementation. 5. IMPLEMENTATION CONTEXT We have implemented a version of the described algorithm and integrated it into a dynamic code generation infrastructure [Kistler 1997; Kistler 1999; Franz 1997; Kistler and Franz 1999] This infrastructure consists of a system that continually profiles all executing code, dynamically generates globally optimized versions of the same code in the background, and then hot swaps the optimized code image in place of the previously executing one. The ....

Franz, M. 1997. Run-Time Code Generation as a Central System Service. In Proceedings of the Sixth Workshop on Hot Topics in Operating Systems (HotOS). 112--117.

....that is an order of magnitude more e#cient than a naive implementation. 5 Dynamic Replacement of Code and Data We have implemented a version of the described algorithm and integrated it into our already existing dynamic code generation infrastructure. This infrastructure consists of a system [Kis96, Fra97] that continually profiles all executing code, dynamically generates globally optimized versions of the same code in the background and then hot swaps the optimized code image in place of the previously executing one. The architecture of this system is modular, allowing it to be readily extended ....

M. Franz. "Run-Time Code Generation as a Central System Service". In The Sixth Workshop on Hot Topics in Operating Systems (HotOS-VI), IEEE Computer Society Press, May 1997.

....the system optimizes applications after execution and hence lacks adaptability as well. In contrast to all of the above work, we are currently developing a system that not only dynamically optimizes programs but also continuously monitors the system s behavior using the above described function [Kis96, Fra97]. In case of substantial changes in the user s behavior, applications are dynamically reoptimized in the background which leads to a much more responsive and adaptive system. 5 Acknowledgement Part of this work is funded by the National Science Foundation under grant CCR 97014000. ....

M. Franz. "Run-Time Code Generation as a Central System Service". In The Sixth Workshop on Hot Topics in Operating Systems (HotOS-VI), IEEE Computer Society Press, pp 112--117, May 1997.

....makes it possible to adapt to the presence of dynamic link libraries as they are loaded, and optimize each one in the context of all the others, not just the ones loaded previously. In the re optimization model, code generation and code optimization become central services of the run time system [Franz 1997]. The advantages are apparent: not only does this solution provide the highest possible code quality, but it also enables every application program to take full advantage of the actual hardware 4 T. Kistler and M. Franz provided that the run time system itself has been targeted toward ....

.... and Schi#man 1984] and SELF [Holzle 1994] systems that focused on the benefits of dynamic optimization in an object oriented environment; Morph , a project developed at Harvard University [Zhang et al. 1997] and the system described by the authors of this paper [Kistler 1997; Kistler and Franz 1997]. Other projects have experimented with optimization at link time rather than at runtime [Wall 1992] At link time, many of the problems described in this paper are non existent. Among them the decision when to optimize, what to optimize, and how to replace code. However, there is also a price to ....

Franz, M. 1997. Run-Time Code Generation as a Central System Service. In Proceedings of the Sixth Workshop on Hot Topics in Operating Systems (HotOS-VI), pp. 112--117. IEEE Computer Society Press.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC