Analysis of Topology-Dependent MPI Performance on Gemini Networks

Analysis of Topology-Dependent MPI Performance on Gemini Networks

Analysis of Topology-Dependent MPI Performance on Gemini Networks Antonio J. Pea, Ralf G. Correa Carvalho, James Dinan, Pavan Balaji, Rajeev Thakur, and William Gropp Motivation Network properties can have a significant impact on application performance BW uses a 3-dimensional Cray Gemini torus featuring anisotropic properties Twice the Y-dimension bandwidth in the X and Z dimensions A Gemini ASIC is shared by two nodes

Task placement considering these properties is highly beneficial EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 2 Outline Background Contributions System: Job Placement and Rank Ordering in BW Network Layout Experimental Evaluation

Basic Micro-benchmarks Collective Communications Stencil Communications Conclusions EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 3 Contributions EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 4 Contributions Characterization of the Gemini anisotropic behavior based in

point-to-point micro-benchmarks Prove Y-wise placement of the dual nodes per network Cartesian point to be highly beneficial Demonstrate potential gains of MPI-network topology matching versus the available node placement EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 5 System: Job Placement / Rank Ordering Network Layout EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 6 Job Placement and Rank Ordering in BW

Cray MPICH follows the node ordering assigned by the job scheduler Ranks are ordered in a zigzag fashion First and last ranks are adjacent Decrease hop count Increase bisection bandwidth Given that: XE6 routers contain two nodes Z links are faster than X links

Every 5th link is crossing a cabinet (slower) 4 x 2 x 8 building blocks Carl Albing, Norm Troullier, Stephen Whalen, Ryan Olson, Joe Glenski, Howard Pritchard, and Hugo Mills. Scalable node allocation for improved performance in regular and anisotropic 3D torus supercomputers. In Recent Advances in the Message Passing Interface, volume 6960 of LNCS, 2011. EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 7 Blue Waters Network Layout X dimension im d Z

n io s en EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 Y dimension Blue Waters Network Layout X dimension Y dimension

EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 Basic Micro-benchmarks Experimental Evaluation EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 10 Point-to-Point Communication (single process) Point-to-point benchmarking

Anisotropic behavior illustrated Communications in the Y direction perform signifficantly lower: links Unidirectional Z links offer much higher TR than Y X and Z: largely different behaviors Latency per hop: ~0.1s Bidirectional

EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 11 Point-to-Point Communication (multiple processes) Internode aggregate transfer rate 2 parallel paths transfer concurrently Optimal node ordering and matching

between MPI ranks and network topo. Collectives saturating links greatly improve performance on Y direction Contiguous nodes in these experiments Double X and Z links become shared Aggregate TR increases for Y Placement of dual nodes/ASIC along Y Extra performance improvement EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 12 Blue Waters Network Layout

X dimension im d Z n io s en EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 Y dimension Collective Communications

Topology matching exploited by row-wise and plane-wise collectives Y direction faster! Row-wise: up to 74% (alltoall) and 54% (allgather) Row-wise MPI_Alltoall Row-wise MPI_Allgather EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 14 Collective Communications Topology matching exploited by row-wise and plane-wise collectives Y direction faster!

Row-wise: up to 74% (alltoall) and 54% (allgather) Plane-wise: up to 59% (alltoall) and 53% (allgather) Plane-wise MPI_Alltoall Plane-wise MPI_Allgather EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 15 Stencil Communications Cray MPICH ignores the reorder

parameter in MPI_Cart_create MPI topo. not matching network 2D & 3D halo exchange (contig. nodes): Plain: Manual ordering X-Y-Z Cart_create: Y-major / Z-Y-X Custom: MPI-network matching 2D

2D: Cart_create worst performance Plain up to 1.4%; Cart_create 4% 3D: Topology matching outperforms MPI-assisted sorting up to 5% Topology matching favors scalability 3D EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 16 Conclusions EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013

17 Conclusions Studied the anisotropic implications of Cray Gemini networks on MPI comms. Characterized this network be means of point-to-point micro-benchmarks Studied the behavior of MPI collectives along the different dimensions / planes

Considering the nodes sharing a network Cartesian coordinate along the Y dimension is highly beneficial, maximizing the use of the network resources Including awareness of the network topology in the MPI library outperforms the available heuristic-based rank ordering Future work: non-contiguous allocations employing existing mapping libraries EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 18

Thank you! EuroMPI 2013 - Madrid (Spain) - 15-18 September 2013 19

Recently Viewed Presentations

  • Ling 390 - Intro to Linguistics - Winter 2005 Class 1 ...

    Ling 390 - Intro to Linguistics - Winter 2005 Class 1 ...

    Calendar Wardhaugh Ch 5 Also will discuss Milroy & Milroy article on Tuesday (notes from me up by Saturday) Quiz 1 on Thursday Particularly "the observer's paradox" - how does Labov resolve this?
  • Epidemiology, Economics, Evolution, and Enlightenment

    Epidemiology, Economics, Evolution, and Enlightenment

    Equity - equity / social justice reflects the distribution of power. Economics - economic resources are fundamental to public health, and their distribution depends on social forces. Evolution - fundamentally, we and our societies are expressions of biology, evolving through...
  • 8.1 Relations and Their Properties

    8.1 Relations and Their Properties

    Let R be a relation from a set A to a set B and S a relation from B to a set C. The composite of. R and S is the relation consisting of ordered pairs (a,c), where aϵA, cϵC...
  • The Bar Of Soap: A multifunctional device with

    The Bar Of Soap: A multifunctional device with

    Daniel Smalley Jeevan Kalanithi Matt Adcock MAS622J 2006 MIT MEDIA LAB Bar of Soap What is the Bar of Soap? A device that can be your phone, your camera, your iPod, your everything. It's not confusing to deal with because:...
  • Advanced manufacturing Lab, Industrial Enginnering Dep ...

    Advanced manufacturing Lab, Industrial Enginnering Dep ...

    Designing a cylindrical pressure vessel: the parameter would be the mean diameter , the thickness, the height . ... Sharif University of Technology. CAD/CAM (21-342), Session #8. Optimization in CAD. Design parameterization . Treatment of constraints.
  • The Three estates - Winston-Salem/Forsyth County Schools

    The Three estates - Winston-Salem/Forsyth County Schools

    Third estate- the rest of French society-the commoners-this, the poorest group, provided most of the country's taxes-these paid for wars, palaces and supported the nobility's lifestyle Directions Today you are going to analyze some political cartoons drawn to represent the...
  • Opus Dei - Middlebury College

    Opus Dei - Middlebury College

    San Francisco: Harper Publishing, 2004. www.odan.org www.opusdeiusa.org Opus Dei The Truth about Opus Dei and The DaVinci Code By: Ellie Molyneux THE SHOWDOWN Dan Brown on Opus Dei "The Vatican Prelature known as Opus Dei is a deeply devout Catholic...
  • Greek Mythology - Bucknell University

    Greek Mythology - Bucknell University

    Conclusions As you can see the ancient Greeks had a very interesting and complex culture which they believed in wholeheartedly. I had a lot of fun researching Greek mythology and I would do it again more extensively anytime. Dionysus Demeter...