Gary Clark Jr. & Junkie XL | Fear the Walking Dead Flight 462 parts 1 to 15 720p.mp4 | January 2017 (1000)
Report from the Inter-Agency Workshop on HPC Resilience John T. Daly

Workshop overview •  3½ day workshop focused on technical challenges of coordinated HPC resilience strategies in the exascale timeframe •  30 participants: system hardware, system software, application developers and users, algorithms and libraries, file systems, I/O and storage, visualization and data analytics •  Highly interactive format –  problem solving teams of not more than 10 persons each –  collaboratively develop a plan and roadmap for implementing resilience

•  Assessed current capabilities, gaps and dependencies for representative applications from “predictive science” (e.g., PDE) and “not predictive science” (e.g., big data) domains •  Created “proof of concept” resilience strategies and an R&D roadmap for a coordinated resilience framework

Goals of the workshop •  Resilience is about keeping the application workload running to a correct solution in a timely and efficient manner in spite of frequent hard (i.e., unrecoverable) and soft (i.e., recoverable) errors •  Demonstrate both the need for and existence of practical resilience strategies that address the future needs of the application via the resources available to the system •  Create a technical strategy and roadmap for addressing resilience in the exascale timeframe –  What are our requirements? What “needs” to work and what can we do without? –  How are we going to meet those requirements?

•  Goals of the workshop DO NOT include… –  Discussion of our current projects or research interests –  Creating new research opportunities –  Figuring out how to get funding for future work

Programmatic considerations •  Resilience strategy should not be minimalist or ad hoc, but neither can we afford a “Cadillac” solution •  If we cannot come up with a credible plan for how we intend to integrate all the pieces then mission is likely to view resources expended on resilience as wasted •  If we ask for the kitchen sink we will end up with nothing; if we short-sell we still end up with nothing  not too small; not too big; just right! •  We need… –  undistracted focus on a “no frills” yet “robust” resilience strategy –  clear delineation between engineering (what we know how to do) and research (what we need to figure out)

•  Apply Occam’s Razor  simplest is best •  I don’t care if it’s an interesting research question; all I care is if I need to figure it out in order to implement a resilience strategy!

Sample “marching orders” •  Working group sessions will be a single assignment divided somewhat arbitrarily into two assignments •  Make as much progress as you can on every part of the assignment; don’t get stuck – identify open issues; move on •  Start with a simple model for resilience; as time, expertise and resources permit then refine your solutions •  Resilience loves anecdotes and anecdotes are all about the edge cases and extremes  use best engineering judgment to focus on middle of the distribution and not the tails •  Focus on high importance items and “low hanging fruit”; do NOT get distracted by “bright shiny objects” (e.g., better ECC) •  Make it in the context of stated (previous day) application requirements –  Convert more hard errors into soft errors –  Provide for reliable and unreliable execution regions –  Empower the application to make some decisions, without negatively impacting the aggregate system workload

Summary of workshop conclusions •  The number of errors, particularly soft errors, occurring on HPC systems will continue to increase •  A right-sized, well-conceived resilience strategy in the exascale timeframe is more cost effective than continuing to rely on ad-hoc resilience solutions •  Must at a minimum provide for a resilience infrastructure that facilitates –  System management of hard errors by effectively “converting” them to soft errors whenever feasible –  Application management of soft errors through interfaces that allow it simple controls over how and when to respond to errors

•  Such a framework is foundational to a deployable and sustainable HPC resilience strategy in the exascale timeframe •  Priorities for R&D in the exascale timeframe: fault characterization, detection, FT algorithms, FT programming models and tools

Priority: Fault characterization •  Reliability will get worse with deeply scaled process technologies creating new modes of failure •  Based on anticipated technology trends, the HPC community needs to develop a useful taxonomy for describing –  the types of faults that future systems are expected to encounter –  their anticipated frequency and impact

Priority: Detection •  In the exascale timeframe, error “recovery” will likely be manageable using known techniques for local checkpointing •  Fault “prediction” is probably too hard a problem to realistically tackle in this timeframe •  The research focus needs to be error “detection” which requires the system and application to work together in a coordinated fashion •  Industry is not going to solve this problem for the HPC community

Priority: FT Algorithms •  Three classes of algorithms were identified a)  those that are embarrassingly fault-tolerant b)  those that are not fault-tolerant but are selfchecking c)  those that are neither fault-tolerant nor selfchecking

•  Most algorithms currently in class (c) could be moved to class (b) or even class (a) by a moderate R&D investment

Priotity: FT programming models •  Resilience would benefit strongly from a programming model that supports some notion of –  transactions in time (e.g., roll-back and recovery) –  transactions in space (e.g., fault containment domains)

•  An uncomplicated, directive-based interface using a handful of assertions (e.g., create persistent memory domains, allocate “reliable” and “unreliable” code regions, etc.) provides most of the necessary interfaces for implementing application fault-tolerance

Priority: Tools •  Resilience lacks a mature, validated test infrastructure to verify the effectiveness of various resilience strategies for keeping the application running in the face of high rates of hard and soft errors •  Fault injection tools are need in particular to simulate all classes of faults •  Models will be required to support the fault testing infrastructure at scale


Apply Occam's Razor simplest is best. • I don't care if it's an interesting research ... FTXS2012_Daly.pdf. FTXS2012_Daly.pdf. Open. Extract. Open with. Sign In.

738KB Sizes 0 Downloads 3 Views

Recommend Documents

No documents