HPC systems keep growing in size to meet the ever-increasing demand for performance and computational resources. Apart from increased performance, large scale systems face two challenges that hinder further growth: energy efficiency and resiliency. At the same time, applications seeking increased performance rely on advanced parallelism for exploiting system resources, which leads to increased pressure on system interconnects. At large system scales, increased communication locality can be beneficial both in terms of application performance and energy consumption. Towards this direction, several studies focus on deriving a mapping of an application's processes to system nodes in a way that communication cost is reduced. A common approach is to express both the application's communication patterns and the system architecture as graphs and then solve the corresponding mapping problem. Apart from communication cost, the completion time of a job can also be affected by node failures. Node failures may result in job abortions, requiring job restarts. In this paper, we address the problem of assigning processes to system resources with the goal of reducing communication cost while also taking into account node failures. The proposed approach is integrated into the Slurm resource manager. Evaluation results show that, in scenarios where few nodes have a low outage probability, the proposed process placement approach achieves a notable decrease in the completion time of batches of MPI jobs. Compared to the default process placement approach in Slurm, the reduction is 18.9% and 31%, respectively for two different MPI applications.
Power efficiency has recently become a major concern in the high-performance computing domain. HPC centers are provisioned by a power bound which impacts execution time. Naturally, a tradeoff arises between power efficiency and computational efficiency. This paper tackles the problem of performance optimization for MPI applications, where a power bound is assumed. The paper exposes a subset of HPC applications that leverage cluster parallelism using MPI, where nodes encounter multiple synchronization points and exhibit inter-node dependency. We abstract this structure into a dependency graph, …
Open source cloud technologies provide a wide range of support for creating customized compute node clusters to schedule tasks and managing resources. In cloud infrastructures such as Jetstream and Chameleon, which are used for scientific research, users receive complete control of the Virtual Machines (VM) that are allocated to them. Importantly, users get root access to the VMs. This provides an opportunity for HPC users to experiment with new resource management technologies such as Apache Mesos that have proven scalability, flexibility, and fault tolerance. To ease the development and depl…
There is a general trend towards solving problems suited to deep learning with more complex deep learning architectures trained on larger training sets. This requires longer compute times and greater data parallelization or model parallelization. Both data and model parallelism have been historically faster in parameter server architectures, but data parallelism is starting to be faster in ring architectures due to algorithmic improvements. In this paper, we analyze the math behind ring architectures and make an informed adaptation of dynamic scheduling to ring architectures. To do so, we form…
The trend for cloud computing has initiated a race towards data centres (DC) of an ever-increasing size. The largest DCs now contain many hundreds of thousands of virtual machine (VM) services. Given the finite lifespan of hardware, such large DCs are subject to frequent hardware failure events that can lead to disruption of service. To counter this, multiple redundant copies of task threads may be distributed around a DC to ensure that individual hardware failures do not cause entire jobs to fail. Here, we present results demonstrating the resilience of different job scheduling algorithms in …
Parallel jobs are different from sequential jobs and require a different type of process management. We present here a process management system for parallel programs such as those written using MPI. A primary goal of the system, which we call MPD (for multipurpose daemon), is to be scalable. By this we mean that startup of interactive parallel jobs comprising thousands of processes is quick, that signals can be quickly delivered to processes, and that stdin, stdout, and stderr are managed intuitively. Our primary target is parallel machines made up of clusters of SMPs, but the system is also …
With the growing constraints on power budget and increasing hardware failure rates, the operation of future exascale systems faces several challenges. Towards this, resource awareness and adaptivity by enabling malleable jobs has been actively researched in the HPC community. Malleable jobs can change their computing resources at runtime and can significantly improve HPC system performance. However, due to the rigid nature of popular parallel programming paradigms such as MPI and lack of support for dynamic resource management in batch systems, malleable jobs have been largely unrealized. In t…
Two-phase I/O is a well-known strategy for implementing collective MPI-IO functions. It redistributes I/O requests among the calling processes into a form that minimizes the file access costs. As modern parallel computers continue to grow into the exascale era, the communication cost of such request redistribution can quickly overwhelm collective I/O performance. This effect has been observed from parallel jobs that run on multiple compute nodes with a high count of MPI processes on each node. To reduce the communication cost, we present a new design for collective I/O by adding an extra commu…
Fault-tolerance has always been an important topic when it comes to running massively parallel programs at scale. Statistically, hardware and software failures are expected to occur more often on systems gathering millions of computing units. Moreover, the larger jobs are, the more computing hours would be wasted by a crash. In this paper, we describe the work done in our MPI runtime to enable both transparent and application-level checkpointing mechanisms. Unlike the MPI 4.0 User-Level Failure Mitigation (ULFM) interface, our work targets solely Checkpoint/Restart and ignores other features s…
C++ advocates exceptions as the preferred way to handle unexpected behaviour of an implementation in the code. This does not integrate well with the error handling of MPI, which more or less always results in program termination in case of MPI failures. In particular, a local C++ exception can currently lead to a deadlock due to unfinished communication requests on remote hosts. At the same time, future MPI implementations are expected to include an API to continue computations even after a hard fault (node loss), i.e. the worst possible unexpected behaviour. In this paper we present an approa…
Recent increased interest in Cloud computing emphasizes the need to find an adequate solution to the load-balancing problem in parallel computing -- efficiently running several jobs concurrently on a cluster of shared computers (nodes). One approach to solve this problem is by preemptive process migration -- the transfer of running processes between nodes. A possible drawback of this approach is the increased overhead between heavily communicating processes. This project presents a solution to this last problem by incorporating the process migration capability of MOSIX into Open-MPI and by red…
Scientific applications have long embraced the MPI as the environment of choice to execute on large distributed systems. The User-Level Failure Mitigation (ULFM) specification extends the MPI standard to address resilience and enable MPI applications to restore their communication capability after a failure. This works builds upon the wide body of experience gained in the field to eliminate a gap between current practice and the ideal, more asynchronous, recovery model in which the fault tolerance activities of multiple components can be carried out simultaneously and overlap. This work propos…
One of the hardest challenges of the current Big Data landscape is the lack of ability to process huge volumes of information in an acceptable time. The goal of this work, is to ascertain if it is useful to use typical Big Data tools to solve High Performance Computing problems, by exploring and comparing a distributed computing framework implemented on a commodity cluster architecture: the experiment will depend on the computational time required using tools such as Apache Spark. This will be compared to "equivalent more traditional" approaches such as using a distributed memory model with MP…
Adaptive workloads can change on--the--fly the configuration of their jobs, in terms of number of processes. In order to carry out these job reconfigurations, we have designed a methodology which enables a job to communicate with the resource manager and, through the runtime, to change its number of MPI ranks. The collaboration between both the workload manager---aware of the queue of jobs and the resource allocation---and the parallel runtime---able to transparently handle the processes and the program data---is crucial for our throughput-aware malleability methodology. Hence, when a job trig…
In this paper, we present a novel and new file-based communication architecture using the local filesystem for large scale parallelization. This new approach eliminates the issues with filesystem overload and resource contention when using the central filesystem for large parallel jobs. The new approach incurs additional overhead due to inter-node message file transfers when both the sending and receiving processes are not on the same node. However, even with this additional overhead cost, its benefits are far greater for the overall cluster operation in addition to the performance enhancement…
The cost of data movement on parallel systems varies greatly with machine architecture, job partition, and nearby jobs. Performance models that accurately capture the cost of data movement provide a tool for analysis, allowing for communication bottlenecks to be pinpointed. Modern heterogeneous architectures yield increased variance in data movement as there are a number of viable paths for inter-GPU communication. In this paper, we present performance models for the various paths of inter-node communication on modern heterogeneous architectures, including the trade-off between GPUDirect commu…
