Subject: C414 JGSI Review Resent-Date: Thu, 30 Sep 1999 22:37:45 -0400 Resent-From: Geoffrey Fox Resent-To: p_gcf@npac.syr.edu Date: Fri, 10 Sep 99 9:23:07 MET DST From: Henri Bal To: gcf@npac.syr.edu Reviewer: Henri Bal, Vrije Universiteit Amsterdam ========================================================================= Review of: More Efficient Serialization and RMI for Java a)Overall Recommendation Accept b)Comments for the authors The paper describes how the performance of Java RMI can be drastically improved using plug-in replacements. This work undoubtedly is very useful to the high-performance Java community, which certainly justifies publication in this special issue. I also have several comments on the paper. The others should take into account that this publication is going to appear in an archival journal on parallel programming in general, whereas the JavaGrande'99 paper was intended for a more specialized (and less formal) forum. More specifically I have the following recommendations: - The paper (in particular the related work section) ignores the large amount of research that has been done outside the Java world on high performance communication protocols. Many of the ideas implemented in KARMI and UKA-serialization have appeared before in other languages and systems (e.g., see Thekkath's work on RPCs, Split-C, other object-oriented languages like Concert or Jade). This does not make the work of the authors less interesting, but they should make more clear which ideas are new and which ideas (from other systems) have been reused for Java. This issue should also be addressed in the conclusions section: summarize which new problems Java RMI introduces besides the ones that are already known and solved by other systems. - Section 3.7 is very JDK specific; it contains too many JDK details for the general parallel computing audience. This part is new compared to the JavaGrande version, but I do not find it very suitable for CP&E. I would write this in a more abstract and concise way, and put the JDK details elsewhere (eg in a report). - The benchmark results lack a discussion of the throughput that is achieved by the various RMI implementations. I would strongly suggest to add a throughput test and to specify the number of memory copies that each RMI implementation makes (since this largely determines throughput on a high-speed network). - The paper sometimes is too informal for a journal. Sentences like "please contribute additional RMI benchmark programs" clearly do not belong in an archival journal. Sore details: - the name of Matt Welsh is misspelled consistently (as Welsch), both in the text and in the references - reference [19] has been published in PPoPP'99 c)Comments to the editor