skip to main content

Explore Research Products in the PAR
It may take a few hours for recently added research products to appear in PAR search results.


Title: Node failure resiliency for Uintah without checkpointing
Summary

The frequency of failures in upcoming exascale supercomputers may well be greater than at present due to many‐core architectures if component failure rates remain unchanged. This potential increase in failure frequency coupled with I/O challenges at exascale may prove problematic for current resiliency approaches such as checkpoint restarting, although the use of fast intermediate memory may help. Algorithm‐based fault tolerance (ABFT) using adaptive mesh refinement (AMR) is one resiliency approach used to address these challenges. For adaptive mesh codes, a coarse mesh version of the solution may be used to restore the fine mesh solution. This paper addresses the implementation of the ABFT approach within the Uintah software framework: both at a software level within Uintah and in the data reconstruction method used for the recovery of lost data. This method has two problems: inaccuracies introduced during the reconstruction propagate forward in time, and the physical consistency of variables, such as positivity or boundedness, may be violated during interpolation. These challenges can be addressed by the combination of two techniques: (1) a fault‐tolerant message passing interface (MPI) implementation to recover from runtime node failures, and (2) high‐order interpolation schemes to preserve the physical solution and reconstruct lost data. The approach considered here uses a “limited essentially nonoscillatory” (LENO) scheme along with AMR to rebuild the lost data without checkpointing using Uintah. Experiments were carried out using a fault‐tolerant MPI‐user‐level failure mitigation to recover from runtime failure and LENO to recover data on patches belonging to failed ranks, while the simulation was continued to the end. Results show that this ABFT approach is up to 10× faster than the traditional checkpointing method. The new interpolation approach is more accurate than linear interpolation and not subject to the overshoots found in other interpolation methods.

  more » « less
PAR ID:
10453870
Author(s) / Creator(s):
 ;  ;  
Publisher / Repository:
Wiley Blackwell (John Wiley & Sons)
Date Published:
Journal Name:
Concurrency and Computation: Practice and Experience
Volume:
31
Issue:
20
ISSN:
1532-0626
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. ; ; ( , Proceedings of the USENIX Conference on File and Storage Technologies (FAST 2021))
    Aguilera, Marcos ; Yadgar, Gala (Ed.)
    Training Deep Neural Networks (DNNs) is a resource-hungry and time-consuming task. During training, the model performs computation at the GPU to learn weights, repeatedly, over several epochs. The learned weights reside in GPU memory, and are occasionally checkpointed (written to persistent storage) for fault-tolerance. Traditionally, model parameters are checkpointed at epoch boundaries; for modern deep networks, an epoch runs for several hours. An interruption to the training job due to preemption, node failure, or process failure, therefore results in the loss of several hours worth of GPU work on recovery. We present CheckFreq, an automatic, fine-grained checkpointing framework that (1) algorithmically determines the checkpointing frequency at the granularity of iterations using systematic online profiling, (2) dynamically tunes checkpointing frequency at runtime to bound the checkpointing overhead using adaptive rate tuning, (3) maintains the training data invariant of using each item in the dataset exactly once per epoch by checkpointing data loader state using a light-weight resumable iterator, and (4) carefully pipelines checkpointing with computation to reduce the checkpoint cost by introducing two-phase checkpointing. Our experiments on a variety of models, storage backends, and GPU generations show that CheckFreq can reduce the recovery time from hours to seconds while bounding the runtime overhead within 3.5%. 
    more » « less
  2. ; ; ( , Proc. of 28th Int. Symp. on High Performance Parallel and Distributed Computing (HPDC'19))
    Transparently checkpointing MPI for fault tolerance and load balancing is a long-standing problem in HPC. The problem has been complicated by the need to provide checkpoint-restart services for all combinations of an MPI implementation over all network interconnects. This work presents MANA (MPI-Agnostic Network-Agnostic transparent checkpointing), a single code base which supports all MPI implementation and interconnect combinations. The agnostic properties imply that one can checkpoint an MPI application under one MPI implementation and perhaps over TCP, and then restart under a second MPI implementation over InfiniBand on a cluster with a different number of CPU cores per node. This technique is based on a novel "split-process" approach, which enables two separate programs to co-exist within a single process with a single address space. This work overcomes the limitations of the two most widely adopted transparent checkpointing solutions, BLCR and DMTCP/InfiniBand, which require separate modifications to each MPI implementation and/or underlying network API. The runtime overhead is found to be insignificant both for checkpoint-restart within a single host, and when comparing a local MPI computation that was migrated to a remote cluster against an ordinary MPI computation running natively on that same remote cluster. 
    more » « less
  3. ; ; ; ; ; ( , Concurrency and Computation: Practice and Experience)
    Summary

    The Exascale Computing Project (ECP) focuses on the development of future exascale‐capable applications. Most ECP applications use the message passing interface (MPI) as their parallel programming model with mini‐apps serving as proxies. This paper explores the explicit usage of MPI in such ECP proxy applications. We empirically analyze 14 proxy applications from the ECP Proxy Apps Suite. We use the MPI profiling interface (PMPI) to collect MPI usage patterns in ECP proxy apps. Our analysis shows that a small subset of features from MPI is commonly used in the proxies of exascale‐capable applications, even when they reference third‐party libraries. This study is intended to provide a better understanding of the use of MPI in current exascale applications. The findings can help focus software investments made for exascale systems in the MPI middleware including optimization, fault‐tolerance, tuning, and hardware‐offload.

     
    more » « less
  4. ; ; ; ( , Proceedings of the ACM on Programming Languages)
    This paper presents a formulation of multiparty session types (MPSTs) for practical fault-tolerant distributed programming. We tackle the challenges faced by session types in the context of distributed systems involving asynchronous and concurrent partial failures – such as supporting dynamic replacement of failed parties and retrying failed protocol segments in an ongoing multiparty session – in the presence of unreliable failure detection. Key to our approach is that we develop a novel model of event-driven concurrency for multiparty sessions. Inspired by real-world practices, it enables us to unify the session-typed handling of regular I/O events with failure handling and the combination of features needed to express practical fault-tolerant protocols. Moreover, the characteristics of our model allow us to prove a global progress property for well-typed processes engaged in multiple concurrent sessions, which does not hold in traditional MPST systems. To demonstrate its practicality, we implement our framework as a toolchain and runtime for Scala, and use it to specify and implement a session-typed version of the cluster management system of the industrial-strength Apache Spark data analytics framework. Our session-typed cluster manager composes with other vanilla Spark components to give a functioning Spark runtime; e.g., it can execute existing third-party Spark applications without code modification. A performance evaluation using the TPC-H benchmark shows our prototype implementation incurs an average overhead below 10%. 
    more » « less
  5. ; ; ; ; ; ( , Proceedings of the Seventh ACM Symposium on Cloud Computing)
    Large-scale distributed systems must be built to anticipate and mitigate a variety of hardware and software failures. In order to build confidence that fault-tolerant systems are correctly implemented, Netflix (and similar enterprises) regularly run failure drills in which faults are deliberately injected in their production system. The combinatorial space of failure scenarios is too large to explore exhaustively. Existing failure testing approaches either randomly explore the space of potential failures randomly or exploit the "hunches" of domain experts to guide the search. Random strategies waste resources testing "uninteresting" faults, while programmer-guided approaches are only as good as human intuition and only scale with human effort. In this paper, we describe how we adapted and implemented a research prototype called lineage-driven fault injection (LDFI) to automate failure testing at Netflix. Along the way, we describe the challenges that arose adapting the LDFI model to the complex and dynamic realities of the Netflix architecture. We show how we implemented the adapted algorithm as a service atop the existing tracing and fault injection infrastructure, and present early results. 
    more » « less
  • Citation Formats
  • MLA
  • APA
  • Chicago
  • BibTeX
  • Save / Share this Record
  • Send to Email