Institut für Informatik
Refine
Has Fulltext
- no (26)
Year of publication
Document Type
- Article (23)
- Monograph/Edited Volume (3)
Is part of the Bibliography
- yes (26)
Keywords
- Cluster Computing (1)
- Cluster computing (1)
- Computational grid (1)
- Green computing (1)
- Load Balancing (1)
- MPI (1)
- Parallel job execution time estimation (1)
- Performance Evaluation (1)
- User submission pattern (1)
- one-sided communication (1)
Institute
Exploring one-sided communication and synchronization on a non-cache-coherent many-core architecture
(2017)
The ongoing many-core design aims at core counts where cache coherence becomes a serious challenge. Therefore, this paper discusses how one-sided communication and the required process synchronization can be realized on a non-cache-coherent many-core CPU. The Intel Single-chip Cloud Computer serves as an exemplary hardware architecture. The presented approach is based on software-managed cache coherence for MPI one-sided communication. The prototype implementation delivers a PUT performance of up to 5 times faster than the default message-based approach and reveals a reduction of the communication costs for the NAS Parallel Benchmarks 3-D fast Fourier Transform by a factor of 5. Further, the paper derives conclusions for future non-cache-coherent architectures.
The Domain Name System belongs to the core services of the Internet infrastructure. Hence, DNS availability and performance is essential for the operation of the Internet and replication as well as load balancing are used for the root and top level name servers.
This paper proposes an architecture for credit based server load balancing (SLB) for DNS. Compared to traditional load balancing algorithms like round robin or least connection, the benefit of credit based SLB is that the load balancer can adapt more easily to heterogeneous load requests and back end server capacities. The challenge of this approach is the definition of a suited credit metric. While this was done before for TCP based services like HTTP, the problem was not solved for UDP based services like DNS.
In the following an approach is presented to define credits also for UDP based services. This UDP/DNS approach is implemented within the credit based SLB implementation salbnet. The presented measurements confirm the benefit of the self-adapting credit based SLB approach. In our experiments, the mean (first) response time dropped significantly compared to weighted round robin (WRR) (from over 4 ms to about 0.6 ms for dynamic pressure relieve (DPR)).
Scheduling performance in computational grid can potentially benefit a lot from accurate execution time estimation for parallel jobs. Most existing approaches for the parallel job execution time estimation, however, require ample past job traces and the explicit correlations between the job execution time and the outer layout parameters such as the consumed processor numbers, the user-estimated execution time and the job ID, which are hard to obtain or reveal. This paper presents and evaluates a novel execution time estimation approach for parallel jobs, the user-behavior clustering for execution time estimation, which can give more accurate execution time estimation for parallel jobs through exploring the job similarity and revealing the user submission patterns. Experiment results show that compared to the state-of-art algorithms, our approach can improve the accuracy of the job execution time estimation up to 5.6 %, meanwhile the time that our approach spends on calculation can be reduced up to 3.8 %.
This paper presents an evaluation of ACPI energy saving modes, and deduces the design and implementation of an energy saving daemon for clusters called cherub. The design of the cherub daemon is modular and extensible. Since the only requirement is a central approach for resource management, cherub is suited for Server Load Balancing (SLB) clusters managed by dispatchers like Linux Virtual Server (LVS), as well as for High Performance Computing (HPC) clusters. Our experimental results show that cherub's scheduling algorithm works well, i.e. it will save energy, if possible, and avoids state-flapping.
Especially for sciences the provision of massive parallel CPU capacity is one of the most attractive features of a grid. A major challenge in a distributed, inherently dynamic grid is fault tolerance. The more resources and components involved, the more complicated and error-prone becomes the system. In a grid with potentially thousands of machines connected to each other the reliability of individual resources cannot be guaranteed.The benefit of the grid is that in case of a failure ail application may be migrated and restarted from a checkpoint file on another site. This approach requires a service infrastructure which handles the necessary activities transparently. In this article, we present Migol, a fault-tolerant and self-healing grid middleware for MPI applications. Migol is based on open standards and extends the services of the Globus toolkit to support the fault tolerance of grid applications.Further, the Migol framework itself is designed with special focus on fault tolerance. For example, Migol eplicates ritical services and uses a ring-based replication protocol to achieve data consistency. (c) 2007 Elsevier B.V. All rights reserved.
Owing to the loose coupling between replicas, the replica-exchange (RE) class of algorithms should be able to benefit greatly from using as many resources as available. However, the ability to effectively use multiple distributed resources to reduce the time to completion remains a challenge at many levels. Additionally, an implementation of a pleasingly distributed algorithm such as replica-exchange, which is independent of infrastructural details, does not exist. This paper proposes an extensible and scalable framework based on Simple API for Grid Applications that provides a general-purpose, opportunistic mechanism to effectively use multiple resources in an infrastructure-independent way. By analysing the requirements of the RE algorithm and the challenges of implementing it on real production systems, we propose a new abstraction (BIGJOB), which forms the basis of the adaptive redistribution and effective scheduling of replicas.
In this paper we present the design and implementation of the Migol brokering framework. Migol is a Grid middleware, which addresses the fault-tolerance of long-running and compute-intensive applications. The framework supports e. g. the automatic and transparent recovery respectively the migration of applications. Another core feature of Migol is the discovery, selection, and allocation of resources using advance reservation. Grid broker systems can significantly benefit from advance reservation. With advance reservation brokers and users can obtain execution guarantees from local resource management systems (LRM) without requiring detailed knowledge of current and future workloads or of the resource owner's policies. Migol's Advance Reservation Service (ARS) provides an adapter layer for reservation capabilities of different LRMs, which is currently not provided by existing Grid middleware platforms. Further, we propose a shortest expected delay (SED) strategy for scheduling of advance reservations within the Job Broker Service. SED needs information about the earliest start time of an application. This is currently not supported by LRMs. We added this feature for PBSPro. Migol depends on Globus and its security infrastructure. Our performance experiments show the substantial overhead of this serviceoriented approach.
Today, InfiniBand is an evolving high speed interconnect technology to build high performance computing clusters, that achieve top 10 rankings in the current top 500 of the worldwide fastest supercomputers. Network interfaces (called host channel adapters) provide transport layer services over connections and datagrams in reliable or unreliable manner. Additionally, InfiniBand supports remote direct memory access (RDMA) primitives that allow for one- sided communication. Using server load balancing together with a high performance cluster makes it possible to build a fast, scalable, and reliable service infrastructure. We have designed and implemented a scalable load balancer for InfiniBand clusters called SLIBNet. Our investigations show that the InfiniBand architecture offers features which perfectly support load balancing. We want to thank the Megware Computer GmbH for providing us an InfiniBand switch to realize a server load balancing testbed.
Seminarband: Sensornetze
(2004)