Efficient deadlock-free routing

Project Information

Discipline

Computer Science (401) 

Orientation

Research 

Efficient deadlock-free routing strategies are crucial to the performance of large-scale computing systems. There are many methods but it remains a challenge to achieve lowest latency and highest bandwidth for irregular or unstructured high-performance networks. The lately published deadlick-free single-source-shortest-path routing algorithm is able to minimize the congestions in the network and is able to avoid deadlocks in the network traffic. We would like to install a patched version of the InfiniBand Open Subnet Manager on India to test the potential performace gain for users and their applications. The users of FG would benefit from the usage of this algorithm on the hpc systems, not only in terms of stability of the network, but also in terms of a speed-up of their MPI programs. The lowered runtime, on the one hand will reduce the costs per project and on the other hand will clear the space for more projects.

We investigated a novel routing strategy based on the single-source-shortest-path routing algorithm and extended it to use virtual channels to guarantee deadlock-freedom. We showed that this algorithm achieves minimal latency and high bandwidth with only a low number of virtual channels and can be implemented in practice. Using FutureGrid hardware these results shall now be backed up.

These tests and results of the performance measurements should encourage the InfinBand consortium to include our patches to the InfiniBand Open Subnet Manager in the official release. The users of FG would benefit from the usage of this algorithm on the hpc systems, not only in terms of stability of the network, but also in terms of a speed-up of their MPI programs.

Project Contact

Project Lead

Thomas William (williath) 

Project Manager

Jens Domke (domke) 

Project Members

Jens Domke  

Resource Requirements

Hardware System

• india (IBM iDataPlex at IU)

 

We would like to use one of the next maintenance windows to deploy our routing algorithm on india's infiniband to test and demonstrate its performance.

As these tests involve changes to the core of the infiniband installation, India needs to be solely available to us during a maintenance for a couple of minutes up to 2h max to install, test and possibly roll back our changes to the InfiniBand Open Subnet Manager

Project Timeline

Submitted

03/02/2012 - 08:24