Modern interconnects offer remote direct memory access (RDMA) features. Yet, most applications rely on explicit message passing for communications albeit their unwanted overheads. The MPI-3.0 standard defines a programming interface for exploiting RDMA networks directly, however, it's scalability and practicability has to be demonstrated in practice. In this work, we develop scalable bufferless protocols that implement the MPI-3.0 specification. Our protocols support scaling to millions of cores with negligible memory consumption while providing highest performance and minimal overheads. To arm programmers, we provide a spectrum of performance models for all critical functions and demonstrate the usability of our library and models with several application studies with up to half a million processes. We show that our design is comparable to, or better than UPC and Fortran Coarrays in terms of latency, bandwidth, and message rate. We also demonstrate application performance improvements with comparable programming complexity.
The MPI standard has long included one-sided communication abstractions through the MPI Remote Memory Access (RMA) interface. Unfortunately, the MPI RMA chapter in the 4.0 version of the MPI standard still contains both well-known and lesser known short-comings for both implementations and users, which lead to potentially non-optimal usage patterns. In this paper, we identify a set of issues and propose ways for applications to better express anticipated usage of RMA routines, allowing the MPI implementation to better adapt to the application's needs. In order to increase the flexibility of th…
Remote-memory-access models, also known as one-sided communication models, are becoming an interesting alternative to traditional two-sided communication models in the field of High Performance Computing. In this paper we extend previous work on an MPI-based, locality-aware remote-memory-access model with a asynchronous progress-engine for non-blocking communication operations. Most previous related work suggests to drive progression on communication through an additional thread within the application process. In contrast, our scheme uses an arbitrary number of dedicated processes to drive asy…
MPI is the most widely used data transfer and communication model in High Performance Computing. The latest version of the standard, MPI-3, allows skilled programmers to exploit all hardware capabilities of the latest and future supercomputing systems. The revised asynchronous remote-memory-access model in combination with the shared-memory window extension, in particular, allow writing code that hides communication latencies and optimizes communication paths according to the locality of data origin and destination. The latter is particularly important for today's multi- and many-core systems.…
Scientific applications often contain large computationally-intensive parallel loops. Loop scheduling techniques aim to achieve load balanced executions of such applications. For distributed-memory systems, existing dynamic loop scheduling (DLS) libraries are typically MPI-based, and employ a master-worker execution model to assign variably-sized chunks of loop iterations. The master-worker execution model may adversely impact performance due to the master-level contention. This work proposes a distributed chunk-calculation approach that does not require the master-worker execution scheme. Mor…
The relaxed semantics and rich functionality of one-sided communication primitives of MPI-3 makes MPI an attractive candidate for the implementation of PGAS models. However, the performance of such implementation suffers from the fact, that current MPI RMA implementations typically have a large overhead when source and target of a communication request share a common, local physical memory. In this paper, we present an optimized PGAS-like runtime system which uses the new MPI-3 shared-memory extensions to serve intra-node communication requests and MPI-3 one-sided communication primitives to s…
Matrix-matrix multiplication is a basic operation in linear algebra and an essential building block for a wide range of algorithms in various scientific fields. Theory and implementation for the dense, square matrix case are well-developed. If matrices are sparse, with application-specific sparsity patterns, the optimal implementation remains an open question. Here, we explore the performance of communication reducing 2.5D algorithms and one-sided MPI communication in the context of linear scaling electronic structure theory. In particular, we extend the DBCSR sparse matrix library, which is t…
Upcoming HPC clusters will feature hybrid memories and storage devices per compute node. In this work, we propose to use the MPI one-sided communication model and MPI windows as unique interface for programming memory and storage. We describe the design and implementation of MPI storage windows, and present its benefits for out-of-core execution, parallel I/O and fault-tolerance. In addition, we explore the integration of heterogeneous window allocations, where memory and storage share a unified virtual address space. When performing large, irregular memory operations, we verify that MPI windo…
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…
This work presents an extension to MPI supporting the one-sided communication model and window allocations in storage. Our design transparently integrates with the current MPI implementations, enabling applications to target MPI windows in storage, memory or both simultaneously, without major modifications. Initial performance results demonstrate that the presented MPI window extension could potentially be helpful for a wide-range of use-cases and with low-overhead.
A Partitioned Global Address Space (PGAS) approach treats a distributed system as if the memory were shared on a global level. Given such a global view on memory, the user may program applications very much like shared memory systems. This greatly simplifies the tasks of developing parallel applications, because no explicit communication has to be specified in the program for data exchange between different computing nodes. In this paper we present DART, a runtime environment, which implements the PGAS paradigm on large-scale high-performance computing clusters. A specific feature of our imple…
In this work, we consider the integration of MPI one-sided communication and non-blocking I/O in HPC-centric MapReduce frameworks. Using a decoupled strategy, we aim to overlap the Map and Reduce phases of the algorithm by allowing processes to communicate and synchronize using solely one-sided operations. Hence, we effectively increase the performance in situations where the workload per process is unexpectedly unbalanced. Using a Word-Count implementation and a large dataset from the Purdue MapReduce Benchmarks Suite (PUMA), we demonstrate that our approach can provide up to 23% performance …
In order to deliver high performance in cloud computing, we generally exploit and leverage RDMA (Remote Direct Memory Access) in networking and NVM (Non-Volatile Memory) in end systems. Due to no involvement of CPU, one-sided RDMA becomes efficient to access the remote memory, and NVM technologies have the strengths of non-volatility, byte-addressability and DRAM-like latency. In order to achieve end-to-end high performance, many efforts aim to synergize one-sided RDMA and NVM. Due to the need to guarantee Remote Data Atomicity (RDA), we have to consume extra network round-trips, remote CPU pa…
Remote Memory Access (RMA), also known as single sided communications, provides a way of accessing the memory of other processes without having to issue explicit message passing style communication calls. Previous studies have concluded that MPI RMA can provide increased performance over traditional MPI Point to Point (P2P) but these are based on synthetic benchmarks. In this work, we replace the existing non-blocking P2P communication calls in the MONC atmospheric model with MPI RMA. We describe our approach in detail and discuss options taken for correctness and performance. Experiments on i…
RDMA is an exciting technology that enables a host to access the memory of a remote host without involving the remote CPU. Prior work shows how to use RDMA to improve the performance of distributed in-memory storage systems. However, RDMA is widely believed to have scalability issues, due to the amount of active protocol state that needs to be cached in the limited NIC cache. These concerns led to several software-based proposals to enhance scalability by trading off performance. In this work, we revisit these trade-offs in light of newer RDMA hardware and propose new guidelines for scaling RD…
TensorFlow has been the most widely adopted Machine/Deep Learning framework. However, little exists in the literature that provides a thorough understanding of the capabilities which TensorFlow offers for the distributed training of large ML/DL models that need computation and communication at scale. Most commonly used distributed training approaches for TF can be categorized as follows: 1) Google Remote Procedure Call (gRPC), 2) gRPC+X: X=(InfiniBand Verbs, Message Passing Interface, and GPUDirect RDMA), and 3) No-gRPC: Baidu Allreduce with MPI, Horovod with MPI, and Horovod with NVIDIA NCCL.…
We present an MPI + OpenACC implementation of the kernel-independent barycentric Lagrange treecode (BLTC) for fast summation of particle interactions on GPUs. The distributed memory parallelization uses recursive coordinate bisection for domain decomposition and MPI remote memory access to build locally essential trees on each rank. The particle interactions are organized into target batch/source cluster interactions which efficiently map onto the GPU; target batching provides an outer level of parallelism, while the direct sum form of the barycentric particle-cluster approximation provides an…
We present a new model for distributed shared memory systems, based on remote data accesses. Such features are offered by network interface cards that allow one-sided operations, remote direct memory access and OS bypass. This model leads to new interpretations of distributed algorithms allowing us to propose an innovative detection technique of race conditions only based on logical clocks. Indeed, the presence of (data) races in a parallel program makes it hard to reason about and is usually considered as a bug.
Fault tolerance for the upcoming exascale generation has long been an area of active research. One of the components of a fault tolerance strategy is checkpointing. Petascale-level checkpointing is demonstrated through a new mechanism for virtualization of the InfiniBand UD (unreliable datagram) mode, and for updating the remote address on each UD-based send, due to lack of a fixed peer. Note that InfiniBand UD is required to support modern MPI implementations. An extrapolation from the current results to future SSD-based storage systems provides evidence that the current approach will remain …
Triangle count and local clustering coefficient are two core metrics for graph analysis. They find broad application in analyses such as community detection and link recommendation. Current state-of-the-art solutions suffer from synchronization overheads or expensive pre-computations needed to distribute the graph, achieving limited scaling capabilities. We propose a fully asynchronous implementation for triangle counting and local clustering coefficient based on 1D partitioning, using remote memory accesses for transferring data and avoid synchronization. Additionally, we show how these algor…
Parallel implementations of linear iterative solvers generally alternate between phases of data exchange and phases of local computation. Increasingly large problem sizes on more heterogeneous systems make load balancing and network layout very challenging tasks. In particular, global communication patterns such as inner products become increasingly limiting at scale. We explore the use of asynchronous communication based on one-sided MPI primitives in a multitude of domain decomposition solvers. In particular, a scalable asynchronous two-level method is presented. We discuss practical issues …
Combining persistent memory (PM) with RDMA is a promising approach to performant replicated distributed key-value stores (KVSs). However, existing replication approaches do not work well when applied to PM KVSs: 1) Using RPC induces software queueing and execution at backups, increasing request latency; 2) Using one-sided RDMA WRITE causes many streams of small PM writes, leading to severe device-level write amplification (DLWA) on PM. In this paper, we propose Rowan, an efficient RDMA abstraction to handle replication writes in PM KVSs; it aggregates concurrent remote writes from different se…
The PGAS model is well suited for executing irregular applications on cluster-based systems, due to its efficient support for short, one-sided messages. However, there are currently two major limitations faced by PGAS applications. The first relates to scalability: despite the availability of APIs that support non-blocking operations in special cases, many PGAS operations on remote locations are synchronous by default, which can lead to long-latency stalls and poor scalability. The second relates to productivity: while it is simpler for the developer to express all communications at a fine-gra…
By providing highly efficient one-sided communication with globally shared memory space, Partitioned Global Address Space (PGAS) has become one of the most promising parallel computing models in high-performance computing (HPC). Meanwhile, FPGA is getting attention as an alternative compute platform for HPC systems with the benefit of custom computing and design flexibility. However, the exploration of PGAS has not been conducted on FPGAs, unlike the traditional message passing interface. This paper proposes FSHMEM, a software/hardware framework that enables the PGAS programming model on FPGAs…
