Section: New Results


Simulation of Parallel Computing Systems

Researchers in the area of distributed computing conduct many of their experiments in simulation. While packet-level simulation is often used to study network protocols, it can be too costly to simulate network communications for large-scale systems and applications. The alternative chosen in SimGrid and a few other simulation frameworks is to simulate the network based on less costly flow-level models. Surprisingly, in the literature, validation of these flow-level models is at best a mere verification for a few simple cases. Consequently, although distributed computing simulators are widely used, their ability to produce scientifically meaningful results is in doubt. In [13] we focus on the validation of state-of-the-art flow-level network models of TCP communication on Wide Area Networks, via comparison to packet-level simulation. While it is straightforward to show cases in which previously proposed models lead to good results, instead we systematically seek cases that lead to invalid results. Careful analysis of these cases reveal fundamental flaws and also suggest improvements. One contribution of this work is that these improvements lead to a new model that, while far from being perfect, improves upon all previously proposed models. A more important contribution, perhaps, is provided by the pitfalls and unexpected behaviors encountered in this work, leading to a number of enlightening lessons. In particular, this work shows that model validation cannot be achieved solely by exhibiting (possibly many) ”good cases.” Confidence in the quality of a model can only be strengthened through an invalidation approach that attempts to prove the model wrong.

The previous results assume steady-state and provide thus a reasonable model when message size is very large. Although, such assumptions may be reasonable when studying grid applications, when simulating HPC applications message sizes are often much smaller and phenomenon like slow-start or how communications and computations overlap have to be accurately modeled. Simulation and modeling for performance prediction and profiling is yet essential for developing and maintaining HPC code that is expected to scale for next-generation exascale systems. In [15] , [34] we describe an implementation of a flow-based hybrid network model that accounts for factors such as network topology and contention, which are commonly ignored by the LogP models. Although, this may seem like a strange choice, we focus on large-scale, Ethernet-connected systems, as these currently compose 37.8% of the TOP500 index, and this share is expected to increase as higher-speed 10 and 100GbE become more available. Furthermore, the European Mont-Blanc project to study exascale computing by developing prototype systems with low-power embedded devices will also use Ethernet-based interconnect [28] . Our model is implemented within SMPI, an open-source MPI implementation that connects real applications to the SimGrid simulation framework. SMPI provides implementations of collective communications based on current versions of both OpenMPI and MPICH. SMPI and SimGrid also provide methods for easing the simulation of large-scale systems, including shadow execution, memory folding, and support for both online and offline (i.e., post-mortem) simulation. We validate our proposed model by comparing traces produced by SMPI with those from real world experiments, as well as with those obtained using other established network models. Our study shows that SMPI has a consistently better predictive power than classical LogP-based models for a wide range of scenarios including both established HPC benchmarks and real applications.

Perfect Simulation

Perfect simulation is a very efficient technique that uses coupling arguments to provide a sample from the stationary distribution of a Markov chain in a finite time without ever computing the distribution. In [7] , we consider Jackson queueing networks (JQN) with finite buffer constraints and analyze the efficiency of sampling from their stationary distribution. In the context of exact sampling, the monotonicity structure of JQNs ensures that such efficiency is of the order of the coupling time (or meeting time) of two extremal sample paths. In the context of approximate sampling, it is given by the mixing time. Under a condition on the drift of the stochastic process underlying a JQN, which we call hyper-stability, in our main result we show that the coupling time is polynomial in both the number of queues and buffer sizes. Then, we use this result to show that the mixing time of JQNs behaves similarly up to a given precision threshold. Our proof relies on a recursive formula relating the coupling times of trajectories that start from network states having 'distance one', and it can be used to analyze the coupling and mixing times of other Markovian networks, provided that they are monotone. An illustrative example is shown in the context of JQNs with blocking mechanisms.

In [35] , we extend the technique to handle situations with infinite space state. We consider open JQN with losses with mixed finite and infinite queues and analyze the efficiency of sampling from their exact stationary distribution. Although the underlying Markov chain may have an infinite state space, we show that perfect sampling is possible. The main idea is to use a JQN with infinite buffers (that has a product form stationary distribution) to bound the number of initial conditions to be considered in the coupling from the past scheme. We also provide bounds on the sampling time of this new perfect sampling algorithm for acyclic or hyperstable networks. These bounds show that the new algorithm is considerably more efficient than existing perfect samplers even in the case where all queues are finite. We illustrate this efficiency through numerical experiments. We also extend our approach to non-monotone networks such as queueing networks with negative customers.