T. Bohman. Shared Memory Computing Architectures for Real-Time Simulation -- Simplicity and Elegance, Technical Report, Systran Corporation, 1994.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....Section 5 presents performance evaluation results. Section 6 presents our conclusions. 2. An Overview of Reflective Memory Systems A Reflective Memory (RM) network provides a limited amount of shared memory across a cluster. There are many reflective memory networks available today. SCRAMNet [11] from Systran, RTNet [12] from VMIC and DEC MemoryChannel [13] are popular reflective memory products. VMMC [14] from Princeton and Merlin and Sesame [15] from Sandia Labs and SUNY, Stony Brook, are research projects involving reflective memory. A detailed survey of existing RM networks is ....

....and O(MN) space respectively for data, where M is the length of the message. The barrier algorithm also requires O(N) space. The amount of shared memory eventually determines how much any algorithm for an RM network can scale. However, today RM NICs are available with up to 8 MB of shared memory [11]. Memory limitation, is therefore not a big problem for medium sized clusters. For instance, a 64 node system, the barrier operation needs only 8 bytes and broadcast and all reduce need only 512 bytes for control. The size of shared memory does not limit the largest message length possible, as ....

T. Bohman. Shared Memory Computing Architectures for Real-Time Simulation -- Simplicity and Elegance, Technical Report, Systran Corporation, 1994.

....and Fiberchannel which provide high bandwidth, but also have high latency due to high software overhead. There is considerable ongoing research for developing low latency networks such as Myrinet [4] and messaging libraries such as FM [12] and U Net [14] SCRAMNet (Shared Common RAM Network) [5, 13] is a replicated non coherent shared memory network. It was initially developed for systems to execute real time applications that require low latency communication, such as aircraft simulators, industrial process control, virtual reality, 1 This research is supported by an NSF SBIR Grant ....

....on top of the BBP. Section 5 desribes performance evaluation results at the BBP and MPI layers. Section 6 describes related work done elsewhere. Section 7 presents our conclusions. 2. SCRAMNet Networking This section briefly describes SCRAMNet networking. It is described in detail in [5, 13]. SCRAMNet consists of network interface cards (NICs) with memory banks connected via a dual ring. The NICs can be plugged into standard I O buses such as PCI and SBus. When a workstation writes to a memory location on its NIC, the write is reflected at the same location on the memory banks on the ....

T. Bohman. Shared-Memory Computing Architectures for Real-Time Simulation-Simplicity and Elegance. Technical report, Systran Corporation, 1994.

....interdisciplinary research project [14] MERKUR isn t the first concept for a bus system interconnecting distributed systems. Many other concepts for so called fieldbusses like PROFIBUS [2, 5] Interbus S [8, 3] CAN [13, 6] ASI [12] SERCOS [15] and the real time network SCRAMNET [4] exist. All these implementations were examined and compared with regard to the important items mentioned above. It s far beyond the scope of this article to discuss this study in all its details here. As a short conclusion, we can establish, that none of the existing fieldbus concepts ....

Tom Bohman. Shared--memory computing architectures for real--time simulation --- simplicity and elegance. Technical description, SYSTRAN Corporation, Dayton, Ohio, 1993.

....any previously sent message is acknowledged. A sender s outstanding unicast messages may have been sent to any other node so acknowledgement by any of these receivers will free up a buffer. 3. 3 Implementation We have initially implemented the BBP on Systran s SCRAMNet reflective memory product [5] with 2MB of memory on each card. Initialization Proper operation relies on each process initializing BBP communication with a unique logical address and the same total number of processes. A fixed size control area of 4096 bytes is allocated to each process. The remaining memory is divided into ....

....by allowing a single copy of message data to be read by multiple receivers. 4 Performance of BBP on SCRAMNet This section describes SCRAMNet, a reflective memory network, and the performance of the BBP implementation on SCRAMNet. 4. 1 Description and characterization of SCRAMNet network SCRAMNet [5] is one type of reflective memory network, providing a limited amount of non coherent shared memory across a cluster. Physically, SCRAMNet is arranged as a ring of up to 256 nodes using fiberoptic or coaxial cable. A fiber optic bypass switch ensures continuity in the ring when not all nodes are ....

T. Bohman. Shared-Memory Computing Architectures for Real-Time SimulationSimplicity and Elegance. Technical report, Systran Corporation, 1994. Matt Jacunski et al.

....any previously sent message is acknowledged. A senders outstanding unicast messages may have been sent to any other node so acknowledgement by any of these receivers will free up a buffer. 3. 3 Implementation We have initially implemented the BBP on Systran s SCRAMNet reflective memory product [5] with 2MB of memory on each card. bbp init( int num procs, int rank ) Allocates and initializes memory for the BBP state structure for process with logical address (rank) of (num procs) This function returns a pointer to a bbp struct which is passed with all other function calls. bbp Send( ....

....by allowing a single copy of message data to be read by multiple receivers. 4 Performance of BBP on SCRAMNet This section describes SCRAMNet, a reflective memory network, and the performance of the BBP implementation on SCRAMNet. 4. 1 Description and characterization of SCRAMNet network SCRAMNet [5] is one type of reflective memory network, providing a limited amount of non coherent shared memory across a cluster. Physically, SCRAMNet is arranged as a ring of up to 256 nodes using fiber optic or coaxial cable. A fiber optic bypass switch ensures continuity in the ring when not all nodes are ....

T. Bohman. Shared-Memory Computing Architectures for Real-Time Simulation-Simplicity and Elegance. Technical report, Systran Corporation, 1994.

....performance in parallel processing, low latency is very critical. As a result considerable effort has been put into developing low latency networking technologies such as Myrinet [5] and messaging software such as Illinois Fast Messages [18] and U Net [23] SCRAMNet (Shared Common RAM Network) [6, 22] is a replicated non coherent shared memory network. It was initially developed for systems to execute real time applications that require low latency communication, such as aircraft simulators, industrial process control, virtual reality, telemetry and robotics. So far, there has not been much ....

....networking technologies. Section 7 describes our ongoing work and related work done elsewhere. Section 8 presents our conclusions. 2 SCRAMNet Networking This section briefly describes the SCRAMNet networking and outlines the advantages for using it in workstation clusters. 2. 1 Overview SCRAMNet [6] is a replicated non coherent shared memory network. It provides a limited amount of non coherent shared memory across the cluster. The network consists of a set of workstations with NICs containing memory banks. Writes to any one memory bank are replicated on all the other NICs on the network. ....

[Article contains additional citation context not shown here]

T. Bohman. Shared-Memory Computing Architectures for Real-Time Simulation-Simplicity and Elegance. Technical report, Systran Corporation, 1994.

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