|
|
SysAdm /
Network SwitchA switch is a networking device that routes packets amongst connected devices. In a cluster, the choice of network equipment, in particular switches, can be critical for performance, especially for MPI jobs. In the UJ cluster, as in many other clusters and batch farms mostly dedicated mostly to natively parallel jobs, we use Gigabit Ethernet, as a cost effective compromise. Also, the adoption of multi-core CPUs and multi-socket servers tends to decrease the networking bandwidth and latency requirements for MPI jobs. Other solutions (Infiniband, Myrinet) could be considered if the user base moves more towards intercommunicating applications, and if funds are available. It should be noted, anyway, that Meraka and CHPC already offer this kind of platforms to the SA research community. At present (Feb 2009) the UJ cluster is using a standard-issue Gigabit switch. As the usage of the cluster grows, we will evaluate the needs and consider acquiring a high-performance switch for the specific use, As of 2008, common wisdom in the HPC community seemed oriented towards the HP ProCurve switches, considered best price/performance and trouble-free. Some suggestions from a friend in Italy:
Other info:
|