Citation Context

...TCP, so these interactions were not taken into account. Without IPR, application execution times were negatively affected by TCP’s wrong estimation of buffers, retransmissions and slow start protocol =-=[6]-=-, that made the whole execution very slow in saturated networks. In contrast, when IPR was applied, jitter was reduced, and this helps TCP to determine its timeout and buffer values, so there were les...

Citation Context

...ion of a multicomputer, and our proposal that mixes two very different simulators, Interconnection Network Simulation and Evaluation Environment [16] (INSEE for short) in charge of the IN, and Simics =-=[7]-=-, used to simulate the compute nodes.sWe also discuss several other approaches to interface these two classes of simulators, and problems that may arise when doing full-system simulation, some due to ...

Citation Context

...rate timing model. However, this setup consumes a huge amount of resources (memory, CPU time) and only allows for the simulation of tens (a few hundreds at most) of interconnected compute nodes. GEMS =-=[8]-=- and SIMFLEX [4] are based on Simics. The Simics environment provides the system simulation, and the other tools provide an accurate timing model which allows doing a highfidelity performance evaluati...

Citation Context

... of starting from scratch, designing a sophisticated Simics module of the target IN, it would be easier for designers to make available tools collaborate. For example, SICOSYS can interface with RSIM =-=[13]-=- to do full-system simulation of shared-memory parallel computers, providing an accurate timing model. However, this setup consumes a huge amount of resources (memory, CPU time) and only allows for th...

Citation Context

...l. However, this setup consumes a huge amount of resources (memory, CPU time) and only allows for the simulation of tens (a few hundreds at most) of interconnected compute nodes. GEMS [8] and SIMFLEX =-=[4]-=- are based on Simics. The Simics environment provides the system simulation, and the other tools provide an accurate timing model which allows doing a highfidelity performance evaluation of systems. T...

Citation Context

...d to a compute node. Fig. 2 represents a model of these routers. Each physical channel of the router is shared by three virtual channels (VCs): an Escape channel (governed by the bubble routing rules =-=[15]-=-), and two adaptive channels. Note that a ring has just one minimal path from source to destination, so packets cannot adapt. Thus, the only difference between the Fig. 2. Model of router simulated by...

Citation Context

... the components that take part in a full-system simulation of a multicomputer, and our proposal that mixes two very different simulators, Interconnection Network Simulation and Evaluation Environment =-=[16]-=- (INSEE for short) in charge of the IN, and Simics [7], used to simulate the compute nodes.sWe also discuss several other approaches to interface these two classes of simulators, and problems that may...

Citation Context

...l-system simulation environments come with default hardware and drivers for that hardware, like Ethernet NICs. Support for other INs such as Myrinet, Infiniband or the torus network of the Bluegene/L =-=[1]-=- is not readily available. These environments provide mechanisms to add new, user-designed hardware modules, than can be integrated into the simulators. If we have an accurate description of a certain...

Citation Context

...ly limit packet injection as soon as the network presents signs of congestion. There are different ways of diagnosing these signs and techniques to avoid congestion, based on global knowledge like in =-=[18]-=-, distributed like RECN [3] or based on information of the local router buffers like LBR [5]. The torus network of the IBM BlueGene/L [1] includes a mechanism that works prioritizing in-transit traffi...

Citation Context

...to integrate sophisticated IN models inside available full-system simulation tools; however, IN simulators are already there as standalone tools (SICOSYS [14], the Chaos Router Simulator [2], FlexSim =-=[17]-=- and many others). Instead of starting from scratch, designing a sophisticated Simics module of the target IN, it would be easier for designers to make available tools collaborate. For example, SICOSY...

Citation Context

...erent ways of diagnosing these signs and techniques to avoid congestion, based on global knowledge like in [18], distributed like RECN [3] or based on information of the local router buffers like LBR =-=[5]-=-. The torus network of the IBM BlueGene/L [1] includes a mechanism that works prioritizing in-transit traffic—we call this IPR (Intransit Priority Restriction). In an initial set of experiments, we ev...

Citation Context

...e simulation of the compute nodes, we use 8 instances of Simics, each one simulating 8 nodes. Each node runs a full Red Hat 7.3 operating system, and can be configured to use some MPI implementations =-=[12]-=-. For this work we chose MPICH [10] because it is widely used and supports several protocol stacks, depending on the underlying IN (Ethernet, Infiniband, Myrinet, …). As we use an Ethernet-like simula...

Citation Context

... soon as the network presents signs of congestion. There are different ways of diagnosing these signs and techniques to avoid congestion, based on global knowledge like in [18], distributed like RECN =-=[3]-=- or based on information of the local router buffers like LBR [5]. The torus network of the IBM BlueGene/L [1] includes a mechanism that works prioritizing in-transit traffic—we call this IPR (Intrans...

Citation Context

...ely parallel processors. It would be possible to integrate sophisticated IN models inside available full-system simulation tools; however, IN simulators are already there as standalone tools (SICOSYS =-=[14]-=-, the Chaos Router Simulator [2], FlexSim [17] and many others). Instead of starting from scratch, designing a sophisticated Simics module of the target IN, it would be easier for designers to make av...