MPWide: A light-weight communication library for wide area message passing and code coupling
0 likes•439 views
MPWide is a communication library designed for wide area message passing, allowing user-space tuning of multiple TCP streams for optimized performance across heterogeneous network configurations. It supports applications such as cosmological n-body simulations and multiscale bloodflow modeling, enabling efficient interconnections between supercomputers with low communication overhead. MPWide is user-friendly, intended for non-admin users, and integrates with existing systems to enhance performance.
MPWide: A light-weight communication library for wide area message passing and code coupling
- 1. MPWide: A communication library for wide area message passing Derek Groen Centre for Computational Science
- 2. Overview The networking landscape Using wide area networks MPWide Example applications Uses for multiscale modelling Questions
- 3. The networking landscape The networks connecting grid sites and supercomputers are highly heterogeneous. − − − Configurations differ at end points. Shared paths vs. Dedicated paths Optical interconnects vs. Regular interconnects. A C B
- 4. The networking landscape Fundamental issue: Networks configurations tend to be node-specific, not path-specific. − What do we do when a node has multiple paths? (most nodes nowadays do) A C B
- 5. Using wide area networks (WANs) Solution 1: Apply a homogeneous configuration for all paths. − Could work for nodes with similar path lengths. − − Not common for WAN communication nodes. Inefficient for the TCP protocol, where the optimal config is dependent on the path length. Requires admin privileges on all end-points. A A A
- 6. Using WANs Solution 2: Adopt a different protocol. − − − − May accomodate heterogeneous configs. New protocol, new list of potential issues. Interplay between protocols on shared networks. Time-consuming and politically heavyweight process. խ ե Ր
- 7. Using WANs Solution 3: User-space tuning through software. − Limited space for tuning. − − Some adjustments require admin rights. Use TCP protocol and existing configurations. No special privileges required. X A Y Y B Z X C Z
- 8. MPWide MPWide is a communication library which allows for user-space tuning of individual paths. For each path it can: − Use 1 or multiple tcp streams. − − Good performance obtained with up to 128 streams/path. Configure different buffer and packet sizes. Apply software-based packet pacing to reduce load. Also improves performance on long networks (Yoshino et al. 2008).
- 9. Example: cosmological N-body One simulation, parallelized across supercomputers. Uses the SUSHI code, which is a cross-site adaptation of GreeM. Models dark matter structure formation over 13.4 billion years. Algorithm: Tree + Particle-mesh. Adaptive load-balancing between sites. →
- 10. Example: cosmological N-body Using 2 to 4 supercomputers simultaneously. − Up to 2048 cores total. MPI within each site. Custom MPWide connections between sites. MPWide Forwarder procs bypass connectivity restrictions. 2048 cores, 3 sites, 7% comm. overhead. 10 Gbps lightpath
- 12. Example: multiscale bloodflow pyNS (1D) coupled to HemeLB (3D). 400.000 time steps, 4000 velocity exchanges − − with 1.2% comm. overhead (512+1 cores, 2298 s), and 5% comm. overhead (2048+1 cores, 907 s.).
- 13. Uses for multiscale modelling Can be used for performance critical cyclic coupling over wide area networks. − High-performance, simple low-level interface. Contains an mpw-cp file transfer client to accelerate file-based couplings. Supports C, C++, Python. Trivial to install and intended for users without administrative privileges. Is being integrated into MUSCLE 2 to improve its coupling performance.
- 14. Thank you! MPWide website: − More on the multiscale bloodflow application: − Groen et al., Interface Focus 3(2), 2013. More on the cosmological N-body application: − http://castle.strw.leidenuniv.nl/software/mpwide.html Groen et al., INFOCOMP 2011, ArXiv:1109.5559. Thanks go out to Steven Rieder, Simon Portegies Zwart, Tomoaki Ishiyama, Keigo Nitadori, Joris Borgdorff, Rupert Nash and the MAPPER consortium as a whole.