Jesus De Oliveira

In this paper we describe KRC, a kernel-level implementation of DSM paradigm using the Release Consistency (RC) Memory Model on the Kerrighed Single System Image
operating system. KRC offers a write-update coherence policy and a Multiple Reader/Multiple Writer (MR/MW) data access algorithm. The main contribution of this work is an adaptive update mechanism which efficiently implements the write-update protocol and the MR/MW data access policy. It consists on the local replication of shared… Show more

In this paper we describe KRC, a kernel-level implementation of DSM paradigm using the Release Consistency (RC) Memory Model on the Kerrighed Single System Image
operating system. KRC offers a write-update coherence policy and a Multiple Reader/Multiple Writer (MR/MW) data access algorithm. The main contribution of this work is an adaptive update mechanism which efficiently implements the write-update protocol and the MR/MW data access policy. It consists on the local replication of shared pages (twin copies) and encoding dynamically the changes made on the original versions, hence, only the differences (diffs) are sent to processes sharing data instead of the full modified pages. We present an experimental evaluation of KRC compared with Kerrighed in terms of total execution time, which shows that the execution of parallel applications on KRC, in particular those with false sharing, has a significant performance improvement over the original Kerrighed platform. We also show that KRC provides linear scalability. Show less

In distributed environments (e.g. grid platform) it is common to find pieces of reusable code distributed among multiple sites. The possibilities of compilation and execution with remote libraries have a great potential to facilitate the integration of pieces of software developed among different organizations. This chapter describes JaDiMa (Java Distributed Machine), a collaborative framework to construct Java applications on grid platforms. JaDiMa automatically manages library repositories to… Show more

In distributed environments (e.g. grid platform) it is common to find pieces of reusable code distributed among multiple sites. The possibilities of compilation and execution with remote libraries have a great potential to facilitate the integration of pieces of software developed among different organizations. This chapter describes JaDiMa (Java Distributed Machine), a collaborative framework to construct Java applications on grid platforms. JaDiMa automatically manages library repositories to allow users to compile and execute applications which use distributed libraries, without keeping these libraries locally. JaDiMa services are implemented as Web Services following the SOA approach; library repositories are modeled as a JXTA P2P network; and semantic annotations of libraries assist developers on the tasks of discovering libraries. We describe an implementation of JaDiMa as part of SUMA/G, a Globus-based grid environment. We show experiences and an empirical evaluation of JaDiMa execution and compilation processes for an application which uses remote libraries for managing graph and network data. Show less

Numerous distributed and transparent data sharing services for parallel processing, specifically targeted at clustered computing architectures have been proposed, aimed at providing a more natural programming environment for parallel application development the more popular but more complex message passing paradigm. Distributed Shared Memory (DSM) systems must address the problem of keeping data consistent when multiple processes running on separate computing nodes accesses and modifies it… Show more

Numerous distributed and transparent data sharing services for parallel processing, specifically targeted at clustered computing architectures have been proposed, aimed at providing a more natural programming environment for parallel application development the more popular but more complex message passing paradigm. Distributed Shared Memory (DSM) systems must address the problem of keeping data consistent when multiple processes running on separate computing nodes accesses and modifies it concurrently, while at the same time, avoiding a high impact on applications execution performance. Relaxed consistency models such as release consistency, entry consistency, and lazy release consistency, have proven a minimal impact on applications performance by reducing the house-keeping communication requirements to maintain data consistency. Assuming that in a correctly designed parallel application, access to shared data items will be guarded by mutual exclusion primitives such as semaphores and barriers, distributed shared data can be temporarily inconsistent, and local modifications can be propagated at the time a process enters or leaves a critical section. This way data consistency communication impact on applications performance is reduced considerably. In this presentation we introduce a transparent and dynamic process reorganisation mechanism aimed at reducing the number of data consistency protocol messages needing to traverse high latency network links. Employing current process migration techniques, such as checkpointing facilities and transparent distributed filesystems, processes that share a common data interest based on the data locality principle, can be accommodated on groups of computing nodes
connected through low latency network links. Through permanent monitoring of network links and processes access patterns to shared data, an suboptimal process organisation can be identified at a given time with a self-organising, distributed algorithm, triggering p.
Show less

Grid technology provides a suitable platform for resource sharing, offering users the possibility of accessing large-scale controlled environments across different organizational boundaries through a virtualized single environment. In order to control accesses to those resources, a unique and global security infrastructure is needed, such as a PKI and Virtual Organizations. Hence, in order to process data in the grid, it must be first uploaded to a suitable resource belonging to the grid. We… Show more

Grid technology provides a suitable platform for resource sharing, offering users the possibility of accessing large-scale controlled environments across different organizational boundaries through a virtualized single environment. In order to control accesses to those resources, a unique and global security infrastructure is needed, such as a PKI and Virtual Organizations. Hence, in order to process data in the grid, it must be first uploaded to a suitable resource belonging to the grid. We propose mechanisms to extend the grid data space boundaries by securely integrating data located in the client local file system or in external repositories. The proposed extensions only take place during execution of an application on the grid, preserving privacy and other security properties. We explain their implementation in suma/g, a middleware built on top of Globus, and show some experiment results. Show less

In this paper we describe KRC-CE, a gLite Single System Image Computing Element. KRC-CE provides to grid users the ability to develop and execute high performing scientific applications which exploit parallelization through a distributed shared memory programming model, instead of the traditional message passing model (MPI). KRC-CE is based on Kerrighed, a kernel-level Single System Image operating system. It implements the Release Memory Consistency Model for System-V shared memory IPC… Show more

In this paper we describe KRC-CE, a gLite Single System Image Computing Element. KRC-CE provides to grid users the ability to develop and execute high performing scientific applications which exploit parallelization through a distributed shared memory programming model, instead of the traditional message passing model (MPI). KRC-CE is based on Kerrighed, a kernel-level Single System Image operating system. It implements the Release Memory Consistency Model for System-V shared memory IPC mechanism with write-update protocol and a Multiple Writer/Multiple Reader data access policy. It provides transparent execution to applications, i.e. no instrumentation/modification is needed to parallel applications that were implemented with shared memory communication to be executed on parallel platforms. KRC-CE ease parallel application programming since the shared memory programming model is often more natural than the message-passing paradigm. KRC-CE clusters implements standard resource discovering mechanisms and job submission and management interfaces to gLite grid components, being presented to the grid users in the same way as the traditional MPI-based clusters. We present some experimental results which shows that applications performance running on KRC-CE clusters is close enough to equivalent implementations using message-passing interprocess communication mechanism such as MPI. Show less

We present a practical experience related to accesing a gLite grid through a VPN Server, instead of using a standard User Interface. In the proposed scheme the User Interface functionallity is located in the user machine and the gLite grid provides a VPN service for accesing grid components. The standard User Interface is replaced by a layer 2 (bridged VPN) User Interface Proxy. The main advantage of using such a scheme is the dramatic improvement in the interaction quality, especially for… Show more

We present a practical experience related to accesing a gLite grid through a VPN Server, instead of using a standard User Interface. In the proposed scheme the User Interface functionallity is located in the user machine and the gLite grid provides a VPN service for accesing grid components. The standard User Interface is replaced by a layer 2 (bridged VPN) User Interface Proxy. The main advantage of using such a scheme is the dramatic improvement in the interaction quality, especially for application developers, because the interaction with a grid through a traditional User Interface is typically limited to secure shell or web sessions. Additionally, the User Interface can be prepackaged in a Virtual Machine, which alleviates the burden of installing a complete User Interface to the user. We show a comparative analysis that involves two different virtual machines (VMWare and qemu) using file transfers through gridFTP. Show less

In this paper we describe the execution model supported by suma/g, a middleware built on top of Globus for execution of Java applications on the grid. This execution model allows a user to launch Java applications that will run on a grid from her machine without requiring this machine to be part of the grid (e.g. a gLite User Interface). Additionally, it allows grid users to regard local file systems, i.e. file systems accessible from their local workstations, as part of the set of file systems… Show more

In this paper we describe the execution model supported by suma/g, a middleware built on top of Globus for execution of Java applications on the grid. This execution model allows a user to launch Java applications that will run on a grid from her machine without requiring this machine to be part of the grid (e.g. a gLite User Interface). Additionally, it allows grid users to regard local file systems, i.e. file systems accessible from their local workstations, as part of the set of file systems accessible within the grid. This eliminates the necessity of performing previous uploads of classes and data files, which helps to meet the grid goal of achieving seamless access to distributed resources. We describe how to implement the services offered by this execution model on Globus-based grids. We compare the use of this execution model with the standard mechanisms for submission of Java jobs in lcg/gLite, a flavor of Globus Toolkit 2. We present experiments showing that this execution model can improve performance for running Java applications on Globus-based grids. Show less

In this paper we introduce and evaluate two prefetching techniques to improve the performance of Java applications executed on the grid. These techniques are experimentally evaluated on two grid environments, by running test applications on two different grid deployment configurations. Our testbed is suma/g, a grid platform specifically targeted at executing Java bytecode on Globus grids. The experimental results show that these techniques can be effective on improving the performance of… Show more

In this paper we introduce and evaluate two prefetching techniques to improve the performance of Java applications executed on the grid. These techniques are experimentally evaluated on two grid environments, by running test applications on two different grid deployment configurations. Our testbed is suma/g, a grid platform specifically targeted at executing Java bytecode on Globus grids. The experimental results show that these techniques can be effective on improving the performance of applications run on the grid, especially for compute intensive scientific applications. Show less

This paper describes JaDiMa (Java Distributed Machine), a collaborative framework to construct high performance distributed Java applications on grid platforms. JaDiMa is a system that automatically manages remote libraries used in a Java application. It leverages on the advantages of portability, modularity, object oriented model and flexibility of Java, while incorporating well known techniques of communication and security. JaDiMa allows users to compile and execute Java applications which… Show more

This paper describes JaDiMa (Java Distributed Machine), a collaborative framework to construct high performance distributed Java applications on grid platforms. JaDiMa is a system that automatically manages remote libraries used in a Java application. It leverages on the advantages of portability, modularity, object oriented model and flexibility of Java, while incorporating well known techniques of communication and security. JaDiMa allows users to compile and execute Java applications which use distributed libraries, without the need of keeping them in the developer and user hosts. The result is a simple and efficient distributed environment upon which applications and data are easily shared and highly portable amongst heterogeneous platforms and multiple users. We describe an implementation of JaDiMa as part of SUMA/G, a Globus-based grid environment. We also show experiences of executing an application, which uses libraries for managing graph and network data, on several scenarios with SUMA/G and JaDiMa. Show less