Entries filed under “Network”

Over at the Energy Sciences Network, Jon Bashor writes that ESnet’s FasterData.es.net is popular online repository of tips and tricks for improving network performance. Maintained by Brian Tierney for the last 15 years or so, the site comprises around 120 pages, about 50% of the page views where on the pages of that focus on tuning Linux hosts for better network performance on network paths above 1 gigabit per second.

There are a few settings that aren’t defaults and if you use them, they can gain you a lot in terms of performance,” said Tierney, cautioning that the same settings will actually downgrade the performance of slower networks or home routers. “The site contains a lot of arcane details that are hard to memorize, so users can go to a page, copy and paste the settings into their systems.”

Read the Full Story.

Intel has been rather quiet about their True Scale InfiniBand products since they acquired QLogic’s IB technology and engineering teams in January 2012. It’s looking like that is starting to change with a series of interviews at The Register and Semi-Accurate that describe how True Scale QDR-80 products measure up to next-generation FDR offerings from Mellanox. To learn more, I caught up with Intel’s Joe Jaworsky last week and got an update on where Intel is heading with fabrics on its journey to Exascale computing.

If you’ll recall, the True Scale architecture replaces traditional Verbs-based MPI libraries with the Performance Scaled Messaging (PSM) layer. Designed specifically with HPC performance in mind, this approach does not require address lookups, which gives you the goodness of lower latency. In large deployments such as those in the Tri-Labs, True Scale has delivered impressive scalability as described in Sandia’s report entitled: Unprecedented Scalability and Performance of the New NNSA tri-lab Linux Capacity Cluster 2, which is part of the SC12 Proceedings.

The performance gains that Sandia users are experiencing with their applications on the new TLCC2 Chama has resulted in many positive feedbacks from happy users…. As pointed out in the introduction, to the scales we have benchmarked on Chama, we are seeing unprecedented performance and scalability of many key Sandia applications.

So how did Intel get True Scale to scale so well? They use something called QDR-80 to double-up 40Gbps QDR ports, thus delivering 80Gbps. While this is a neat trick in itself, Intel has paired this with lower pricing than FDR alternatives. The net-net of this is that Intel is offering low latency interconnects for less money, a persuasive message indeed for supercomputer buyers who would love to take that extra cash and use it to buy more cores.

Watch this space as Intel continues to beef up its portfolio on its journey to Exascale computing. Compute is just one element of a performing, balanced architecture. As Intel brings fabrics (Cray Aries, True Scale, and Fulcurm) and storage (Lustre expertise and SSDs) to bear, the company is building a formidable platform for the future of HPC.

Over at the HP Blog, Steve Barry writes that the days of Blade servers being hampered by the limitations of top-of-rack switches may be over thanks to new technology from Mellanox. Designed specifically for customers that demand performance and raw bandwidth the Mellanox SX1018HP blade switch provides up to sixteen 40Gb server downlinks and up to eighteen 40Gb network uplinks for an amazing 1.3Tb/s of throughput.

For customers currently running Infiniband networks, the appeal of being able to collapse their data requirements onto a single network has always been tempered by the lack of support for Remote Direct Memory Access (RDMA) on Ethernet networks. Again, HP and Mellanox lead the way in blade switches. The SX1018HP supports RDMA over Converged Ethernet (RoCE) allowing those RDMA tuned applications to work across both Infiniband and Ethernet networks. When coupled with the recently announced HP544M 40Gb Ethernet/FDR Infiniband adapter, customers can now support RDMA end to end on either network and begin the migration to a single Ethernet infrastructure.

Read the Full Story.

Over at the Boundary Blog, Steven Strowes writes that understanding network delay is key to understanding some important aspects of your network performance.

This post describes how Boundary uses a well-known TCP mechanism to calculate round-trip times (RTTs) between any two hosts by passively monitoring TCP traffic flows, i.e., without actively launching ICMP echo requests (pings). The post is primarily an overview of this one aspect of TCP monitoring, it also outlines the mechanism we are using, and demonstrates its correctness.

Read the Full Story.