Scott Pakin, Vijay Karamcheti, and Andrew A. Chien, "Fast Messages: Efficient, portable communication for workstation clusters and MPPs," IEEE Concurrency, vol. 5, no. 2, pp. 60 73, Apr. qune 1997, Available: http://www- csag. ucsd. edu/ papers/fra-pdt.ps.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....by two independent networks. One is a switched Ethernet which serves as a control network, using the conventional TCP IP protocol and a reliable multicast protocol developed locally [8] The other network is a 1. 28 Gb s Myrinet dedicated to the users applications communication using MPI over FM [11]. The ParPar software is based on the Unix BSDI system, and runs at user level. It includes daemons which run on the host (masterd) and on the nodes (noded) and graphical user interfaces for each running job (job rep) In addition there are the processes (sprocs) which make up the parallel jobs ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast messages: efficient, portable communication for workstation clusters and MPPs ". IEEE Concurrency 5(2), pp. 60--73, Apr-Jun 1997.

....connected by a 10MB switched Ethernet that serves for control functions, and a 1.28GB Myrinet [1] for data communications. The Myrinet network interface cards have a LANai 4.3 processor and 512 KB RAM. Data communications use a modified version of the FM 2. 0 library from the University of Illinois [10] (more in Section 3) The software is a set of daemons: a master daemon, masterd, is run on one machine which is considered the host (or manager) of the cluster, and is not used by the user applications. Every other node runs a node daemon, noded, which manages the processes on this node. These ....

....buffers. Finally the nodeds signal the Myrinet card to return to normal communication status. This is explained in detail in Section 3.2. 2.2. The Fast Messages (FM) Library The FM system was developed at the University of Illinois as part of the High Performance Virtual Machine (HPVM) project [10]. The system offers high speed communication over Myricom s Myrinet SAN (System Area Network) It is composed of the following components: A library that is linked to user applications and contains an initialization routine and the basic routines for sending and receiving messages. A control ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast messages: efficient, portable communication for workstation clusters and MPPs ". IEEE Concurrency 5(2), pp. 60--73, Apr-Jun 1997.

....performance is the high software overhead[4, 13, 14] which mainly results from switching context and copying data between the user and the kernel spaces. To overcome these problems, many light weight protocols have been proposed to move the protocol stacks from the kernel to the user space[2, 9, 16, 19, 20]. One of these protocols is Virtual Interface Architecture(VIA) 9] which was jointly proposed by Intel, Compaq, and Microsoft. The VIA specification describes a network architecture for protected user level zero copy communication. VIA gives an illusion to a user process as if the process has a ....

S. Pakin, V. Karamcheti, and A. A. Chien. "Fast Messages: Efficient, Portable Communication for Workstation Clusters and MPPs". IEEE Concurrency, 5(2):60--73, April-June 1997.

....networking mechanisms designed to minimize communication overhead. The VIA specification [VIA97] was developed by industry leaders as a proposed standard for high performance communication in distributed systems. It combines the principles of Active Messages and many other user level networks [Pak97, Eic95, Pry98, Dub96, Gil96, Dru94] as well as traditional network architectures into a combined software hardware design. Split C has been implemented on many MPPs, as well as the Berkeley NOW (a cluster of 100 UltraSPARC 1s) This paper describes several new implementations of the Split C communications layer targeted for the ....

S. Pakin, V. Karamcheti, A. A. Chien. "Fast messages: efficient, portable communication for workstation clusters and MPPs." IEEE Concurrency, vol.5, (no.2), AprilJune 1997. p.60-72.

....either all the sprocs in a job are running on distinct processors, or none are running. The first issue in the implementation of gang scheduling is how to map sprocs to nodes. We use an extension of the Distributed Hierarchical Control framework, which defines a binary tree of control points [11]. The extension supports non powerof two systems, and builds the control tree incrementally as additional nodeds join the system (Fig. 1.8) This scheme nevertheless tends to pack jobs using groups of processors that are powers of two, which improves the ability to perform alternative scheduling. ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast messages: efficient, portable communication for workstation clusters and MPPs". IEEE Concurrency 5(2), pp. 60--73, Apr-Jun 1997.

....with two or three dual switches. Left: the abstract topology. Right: the Myrinet implementation. 1. 2 Fast Messages Fast Messages (FM) is a high performance communication library for Myrinet, operating at the user level (that is, sending messages does not involve the operating system kernel) [24]. To integrate FM with the ParPar system, the daemons used by FM to set up its environment (GRM and CM) were replaced by functions in the ParPar daemons. This was relatively straightforward, as the required data such as job ID and logical node numbers exists in the ParPar system anyway ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast messages: efficient, portable communication for workstation clusters and MPPs". IEEE Concurrency 5(2), pp. 60--73, Apr-Jun 1997.

....reliability, simplicity and lower cost. With the supports of such high performance interconnection networks, multiple SHV servers can be connected to form a powerful supercomputing environment. In the past, various fast messaging mechanisms for clusters have been proposed, such as AM [4] FM [5], UNet [6] VMMC [13] and BIP [7] These mechanisms have been ported on Fast Ethernet, ATM or Myrinet. Recently several prototype cluster communication systems using Gigabit networking have been built. For example, Berkeley s Linux VIA [9] is a high performance implementation of the Virtual ....

S. Pakin, V. Karamcheti, A. A. Chien. "Fast Messages: Efficient, Portable Communication for Workstation Clusters and MPPs", IEEE Concurrency, vol.5, (no.2), April-June 1997, pp.60-72.

....passes through it, using the conventional TCP IP protocol and a reliable multi cast protocol developed locally [5] The other network is a 1. 28 GB S Myrinet dedicated to the users applications communication (this network is called the data network) The communication is done using MPI over FM [17]. The data network connects only the 16 nodes (user processes don t run on the host) 3.1.2 Software structure The ParPar software is based on the Unix BSDI system, and runs at user level. The software components of the ParPar system includes daemons which run on the host and on the nodes, and ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast messages: efficient, portable communication for workstation clusters and MPPs". IEEE Concurrency 5(2), pp. 60--73, Apr-Jun 1997.

....of order of magnitude improvements in raw network speed to applications. Several research projects have proposed and implemented fast networking protocols and interfaces that allow user level access to high speed networking devices. Examples include Cornell U Net [V95] Illinois Fast Messages (FM) [P97], Princeton Virtual Memory Mapped Communication (VMMC) B94] etc. Another trend in distributed computing is the increasing popularity of server clusters. By connecting a number of relatively inexpensive machines with high speed System Area Networks (SANs) such configurations offer a ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast Messages: Efficient, Portable Communication for Workstation Clusters and MPPs," IEEE Concurrency, 5(2):60-73, 1997.

....and the Sun port was not robust. Thus, a major contribution of this paper is to advocate a framework for writing blockstructured asynchronous collective communication algorithms. This framework would serve as middleware sitting atop communication APIs like MPI, but could use other APIs as well [32]. The Data Mover resembles an MPI persistent communication object. However, KeLP provides inspector executor analysis, which is particularly useful in irregular problems, and it also provides first class support for multidimensional arrays, via user defined metadata and a geometric region ....

S.Pakin, V.Karamcheti, and A.A. Chien, "Fast Messages: Efficient Portable Communication for Workstation Clusters and MPPs," IEEE Concurrency, Vol. 5, No. 2, 1997, pp. 60--72.

....inappropriateness of the mechanism. Thus, a major contribution of this paper is to advocate a framework for writing blockstructured asynchronous collective communication algorithms. This framework would serve as middleware sitting atop communication APIs like MPI, but could use other APIs as well [32]. MPI s nearest equivalent operation to the KeLP Mover is Alltoallw. However, MPI does not mandate an asynchronous version of Alltoallw, introducing the problem of nonportability. In addition, MPI s primitive does not support many to one mappings of data to processors needed to handle load ....

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast messages: efficient portable communication for workstation clusters and MPPs," IEEE Concurrency, vol. 5, no. 2, pp. 60--72, 1997.

.... have been shown not to be able to deliver the unprecedented level of performance made available [8] A number of high performance communication projects have been started to reduce the gap, like U Net [12] and SHRIMP [3] This has also been the objective of the Fast Messages (FM) project [7, 9, 11, 10]. In such project a highly optimized, low latency messaging layer providing a virtual interface to the hardware was developed for the Myrinet network [4] The design of FM addresses the critical issues found in building a low level messaging layer: division of labor between the host and the ....

S. Pakin, V. Karamcheti, and A. Chien, "Fast Messages: Efficient, Portable Communication for Workstation Clusters and MPPs", IEEE Concurrency, vol. 5(2), April-June, 1997, pp. 60--73.

....Fast Fourier Transforms, and blocked algorithms for dense linear algebra. We assume that SMP nodes communicate via message passing, and that message passing incurs high software overheads. We do not assume fast sharedmemory access between nodes (e.g. 94, 77, 115, 133, 28] low overhead messages [136, 107, 33] or an efficient software based distributed shared memory system [88, 40, 60] We make no claims regarding the utility, performance, or economic feasibility of these specialized solutions. Indeed, tightly coupled scalable shared memory systems may well pervade the high performance computer market ....

S. Pakin, V. Karamcheti, and A. A. Chien. Fast messages: efficient portable communication for workstation clusters and MPPs. IEEE Concurrency, 5(2):60--72, 1997.

....A communication subsystem provides low level message passing mechanisms to support DSMs. Efficiency of the subsystem impacts directly on the supported DSM. An intuitive design strategy is to bypass the bulky communication software, e.g. TCP IP, and replace it with a highly optimized subsystem [7, 23]. Unfortunately, since this approach usually requires a higher integration with the lowest level hardware, portability of the resultant system is a problem. If portability is a major concern, then it is a good design choice to use existing, widely accepted communication subsystems, while leaving ....

S. Pakin, V. Karamcheti, and A. Chien, "Fast Messages: Efficient, Portable Communication for Workstation Clusters and Massively-Parallel Processors," IEEE Parallel and Distributed Technology , 1997 (To appear).

....HEN gigabit networks, such as Myrinet [1] Giganet [2] and Gigabit Ethernet [3] started to become commonplace on workstation clusters, the communication performance bottleneck shifted from network hardware to the messaging software. A number of highly optimized messaging layers Fast Messages [4], Active Messages [5] and U Net [6] to name a few arose to bridge the gap between the performance achievable by the hardware and that observed by applications. As a result of these efforts, applications can observe point to point communication performance near that achievable by the raw net ....

Scott Pakin, Vijay Karamcheti, and Andrew A. Chien, "Fast Messages: Efficient, portable communication for workstation clusters and MPPs," IEEE Concurrency, vol. 5, no. 2, pp. 60 73, Apr. qune 1997, Available: http://www- csag. ucsd. edu/ papers/fra-pdt.ps.

No context found.

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast Messages: efficient, portable communication for workstation clusters and MPPs," IEEE Concurrency, vol. 5, p. 90, Apr. 1997.

No context found.

S. Pakin, V. Karamcheti, and A. A. Chien, "Fast Messages: efficient, portable communication for workstation clusters and MPPs," IEEE Concurrency, vol. 5, p. 90, Apr. 1997.

No context found.

S. Pakin, V. Karamcheti, and A. Chien, "Fast Messages: Efficient, portable communication for workstation clusters and MPPs", IEEE Concurrency, Vol. 5(2), Spring 1997, pages 60--- 72.

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