Michael Philippsen, Bernhard Haumacher, and Christian Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....performs many computations that are redundant in the context of highperformance computing. For example, when an object is serialized, the object s class information is serialized as well. Other researchers have worked on improving JDKserialization. Good examples include UKA serialization [28], the Manta project [19] and Jaguar [44] The first two approaches apply a new wire format and require explicit serialization and de serialization methods for each class. The third approach uses pre serialized objects whose memory layout is already in a serialized form. Details of how our ....

....Java Sockets. Other performance results can be found in [40] Table 5. Comparison of standard and CO 2 P 3 S serialization (milliseconds) Objects Strings Integers Standard 31.4 4.2 5. 6 16 9 13 We also ran additional tests to compare our enhancements directly to UKA serialization work [28] and to standard JDKserialization outside of the CO 2 P 3 S environment. Four kinds of data were used in the tests: an array with 1000 TransportableTree elements, one TransportableTree object, an array with 1000 TestClass elements and one TestClass object. The TransportableTree class implements ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495-518, May 2000.

....mechanism described in Section 2.2, the ports can be configured to use Sun serialization or Ibis serialization. Table 5 shows the latencies and throughputs that are achieved by several RMI implementations on our hardware using the IBM JIT. The Sun RMI implementation only runs over TCP. KaRMI [27] is an optimized serialization and RMI implementation. On TCP, KaRMI is implemented in pure Java, as is Ibis RMI on TCP. There also is a KaRMI version that uses native code to interface with GM, which makes KaRMI a good candidate for performance comparison with Ibis RMI, both on TCP and GM. As ....

....for the Java MPI system in [13] the latency for calling MPI from Java is 119 s higher than calling MPI from C (346 versus 227 s, measured on an SP2) RMI performance is studied in several other papers. KaRMI is an improved RMI and serialization package designed to improve RMI performance [27]. The performance of Ibis RMI is better than that of KaRMI (see Table 5 in Section 4) The main reason is that Ibis generates serialization code instead of using run time type inspection. Also, Ibis exploits features of the underlying communication layer, and allows a zero copy implementation by ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

....code generated by the Manta compiler. 6. RELATED WORK Java is increasingly recognized as a suitable platform for highperformance computing. The driving force in high performance Java is the Java Grande Forum (www.javagrande.org) Many research projects investigate parallel programming in Java [7, 8, 12, 22]. Efficient communication mechanisms are a vital building block for high performance Java [15] One approach to sharing objects between parallel processes is to use either object caching [2, 25] or object replication [16] However, neither replication nor caching of objects provides the ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

....to achieve interoperability between Java and HPC , a version of C with library extensions for parallel computing [16] Since Nexus is a binary protocol, NexusRMI achieves efficient performance when communicating with HPC . 6] To gain significant improvement in performance, Philippsen et al. [28] built drop in replacements in their RMI implementations. Thiruvathukal et al. 31] used explicit methods to serialize and deserialize each object s internal state. Their implementation of Java RMI provides unique features like dynamically obtaining the interface of remote objects. In [32] ....

Michael Philippsen, Bernhard Haumacher, and Christian Nester. More efficient serialization and RMI for Java. to appear in Concurrency: Practice and Experience, 2000.

....basis for writing parallel programs. RMI uses a familiar abstraction (object invocation) integrated in a clean way in Java s object oriented programming model. For example, almost any data structure can be passed as argument or return value in an RMI. Also, RMI can be implemented efficiently [21, 25] and it can be extended seamlessly with support for object replication [20] A disadvantage of RMI, however, is that it only supports communication between two parties, a client and a server. Experience with other parallel languages has shown that many applications also require communication ....

....rendered pixels to CPU 0 by individual messages. The speedups achieved by mpiJava and by CCJ are almost identical. 6 Related work The driving force in high performance Java is the Java Grande Forum (www.javagrande.org) There are also many other research projects for parallel programming in Java [1, 6, 7, 14, 16, 25]. Most of these systems, however, do not support collective communication. Taco [24] is a C template library that implements collective operations, however without exploiting MPI s concept of collective invocation by the participating processes. JavaNOW [27] implements some of MPI s collective ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

....Discussion and empirical results are presented in the context of our public benchmark suite. 1 Introduction To date much of the study of Java for high performance computation can be grouped into one or more of a few broad categories: lifting language constraints [2] optimizing communication [6, 8, 9], extending numerical capability [3] and various forms of native compilation [7 9] These and related contributions demonstrate clearly that Java platform can be used for numerical work, but by and large they retain a research orientation in that performance gains are obtained through highly ....

M. Phillipsen, B. Haumacher, C. Nester, "More Efficient Serialization and RMI for Java", Concurrency: Practice and Experience, 12(7):495-518, May 2000

....17: Speedup for the QR application 5. RELATED WORK Java is increasingly recognized as a suitable platform for highperformance computing. The driving force in high performance Java is the Java Grande Forum (www.javagrande.org) Many research projects investigate parallel programming in Java [1, 9, 10, 16, 19, 28]. Efficient communication mechanisms have been recognized as a vital building block for high performance Java [21] Most existing systems, however, do not support efficient group communication. In our previous work on parallel Java, we implemented several applications based on RMI and RepMI ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

....basis for writing parallel programs. RMI uses a familiar abstraction (object invocation) integrated in a clean way in Java s object oriented programming model. For example, almost any data structure can be passed as argument or return value in an RMI. Also, RMI can be implemented efficiently [9, 12] and it can be extended seamlessly with support for object replication [8] A disadvantage of RMI, however, is that it only supports communication between two parties, a client and a server. Experience with other parallel languages has shown that many applications also require communication ....

....only 11.4 of the RMI version. The RMI version of LEQ is 72 larger than the CCJ version. This is is caused by the communication code, which makes up 67 of the RMI version, but only 29 of the CCJ version. 5 Related work There are many other research projects for parallel programming in Java [2, 3, 12]. Most of them, however, do not support collective communication. Taco [11] is a C template library that implements collective operations. JavaNOW [13] implements some of MPI s collective operations on top of a Linda like entity space; however, performance is not an issue. In our previous work ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495-518, May 2000. http://wwwipd.ira.uka.de/JavaParty/.

No context found.

Michael Philippsen, Bernhard Haumacher, and Christian Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, to appear, 2000.

....on remote objects do not work properly. Thirdly, the thread interruption mechanism for threads executing a remote call is broken. These problems make multi threaded distributed programming complicated and error prone. We present a two level solution: On the library level, we extend KaRMI [16], a fast replacement for RMI, with global thread identities for eliminating problems with monitor reentry. Problem with synchronization on remote objects are solved with a facility for remote monitor acquisition. Our interrupt forwarding mechanism enables the application to get full control over ....

....next inner remote method is called. Only the representative of the head segment is currently runnable and represents the distributed thread s point of execution. 2.1 KaRMI, a drop in replacement for RMI RMI is designed for wide area, high latency networks and uses a slow object serialization. In [16] we introduced KaRMI, a much faster drop in RMI and an efficient drop in serialization designed and implemented completely in Java without any native code. Moreover, this redesigned RMI supports non TCP IP communication networks, for instance high performance networks such as Myrinet. KaRMI keeps ....

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, May 2000.

....a remote method invocation in JavaParty, although fully implemented in Java, takes about 40 s on a cluster of dual Pentium processors connected by Myrinet. The central ideas of the optimization will be highlighted in the next two sub sections. 4. 1 Fast UKA Serialization The UKA serialization [15] can be used instead of the official serialization (and as a supplement to it) Table 1 shows the effect of the UKA serialization for several different types of objects. For an object with 32 int values, instead of 51 239=290 s for serialization and de serialization with the standard JDK, 10 ....

....at the University of Karlsruhe. Figure 3 shows that, for benchmark programs, 91 of the time can be saved, if the UKA serialization, the high speed RMI (KaRMI) and the faster communication hardware are all used. A number of deficiencies of the official RMI have been solved by our high speed KaRMI [15], which can be used instead for grande applications. The list of these improvements include: Java Communications for Large Scale Parallel Computing 11 Runtume improvements compared to RMI JDK serialization over Ethernet 0,00 25,00 50,00 75,00 100,00 left: KaRMI UKA serialization over ....

M. Philippsen, B. Haumacher, and C. Nester. More Efficient Serialization and RMI for Java. Concurrency: Practice and Experience, vol. 12(7):495--518, 2000.

....slow for high performance applications that run on low latency networks, for example on a closely connected cluster of workstations. Several projects are underway to improve the performance of RMI, for example the Manta project [7] and the JavaParty project (http: wwwipd.ira.uka.de JavaParty ) [10]. In addition to simply improve the implementation for better speed, the main optimization ideas are first, to use pre compiled marshalling routines instead of the ones that are generated at run time by means of dynamic type introspection. A second idea is, to use only as little bandwidth for type ....

M. Philippsen, B. Haumacher, and C. Nester. More Efficient Serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, May 2000.

No context found.

Michael Philippsen, Bernhard Haumacher, and Christian Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. pp. 495--518, May 2000.

No context found.

Michael Philippsen, Bernhard Haumacher, and Christian Nester, "More Efficient Serialization and RMI for Java", Concurrency: Practice & Experience, 12(7):495-518, May 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7), pp. 495--518, May 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495-518, May 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

No context found.

Michael Philippsen, Bernhard Haumacher, and Christian Nester, "More Efficient Serialization and RMI for Java", Concurrency: Practice & Experience, 12(7):495-518, May 2000.

No context found.

Michael Philippsen, Bernhard Haumacher, and Christian Nester, "More Efficient Serialization and RMI for Java", Concurrency: Practice & Experience, 12(7):495-518, May 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, May 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

No context found.

M. Philippsen, B. Haumacher, and C. Nester. More efficient serialization and RMI for Java. Concurrency: Practice and Experience, 12(7):495--518, 2000.

First 50 documents

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