Top 500 Interconnect Trends

Published twice a year and publicly available at www.top500.org, the TOP500 supercomputing list ranks the world’s most powerful computer systems according to the Linpack benchmark rating system.

Taken from Nvidia Networking

Summary of Findings for Nvidia Networking.

  • NVIDIA GPU or Network (InfiniBand, Ethernet) accelerate 342 systems or 68% of overall TOP500 systems
  • InfiniBand accelerates seven of the top ten supercomputers in the world
  • NVIDIA BlueField DPU and HDR InfiniBand Networking accelerate the world’s 1st academic cloud-native supercomputer at Cambridge University
  • NVIDIA InfiniBand and Ethernet networking solutions connect 318 systems or 64% of overall TOP500 platforms
  • InfiniBand accelerates 170 systems, 21% growth compared to June 2020 TOP500 list
  • InfiniBand accelerates #1, #2 supercomputers in the US, #1 in China, #1, #2 and #3 in Europe
  • NVIDIA 25 gigabit and faster Ethernet solutions connect 62% of total Ethernet systems

Rapid Growth in HPC Storage

The Article is taken from On-Prem No Longer Centre Stage for Broader HPC Storage

AI/ML, more sophisticated analytics, and larger-scale HPC problems all bode well for the on-prem storage market in high performance computing (HPC) and are an even bigger boon for cloud storage vendors.

Nossokoff points to several shifts in the storage industry and among the top supercomputing sites, particularly in the U.S. that reflect changing priorities with storage technologies, especially with the mixed file problems AI/ML introduce into the traditional HPC storage hierarchy. “We’re seeing a focus on raw sequential large block performance in terms of TB/s, high-throughput metadata and random small-block IOPS performance, cost-effective capacity for increasingly large datasets in all HPC workloads, and work to add intelligent placement of data so it’s where it needs to be.”

In addition to keeping pace with the storage tweaks to suit AI/ML as well as traditional HPC, there have been shifts in the vendor ecosystem this year as well. These will likely have an impact on what some of the largest HPC sites do over the coming years as they build and deploy their first exascale machines. Persistent memory is becoming more common, companies like Samsung are moving from NVMe to CXL, which is an indication of where that might fit in the future HPC storage and memory stack. Companies like Vast Data, which were once seen as an up and coming player in the on-prem storage hardware space for HPC transformed into a software company, Nossokoff says.

On-Prem No Longer Centre Stage for Broader HPC Storage – NextPlatform

UDP Tuning to maximise performance

There is a interesting article how your UDP traffic can maximise performance with a few tweak. The article is taken from UDP Tuning

The most important factors as mentioned in the article is

  • Use jumbo frames: performance will be 4-5 times better using 9K MTUs
  • packet size: best performance is MTU size minus packet header size. For example for a 9000Byte MTU, use 8972 for IPV4, and 8952 for IPV6.
  • socket buffer size: For UDP, buffer size is not related to RTT the way TCP is, but the defaults are still not large enough. Setting the socket buffer to 4M seems to help a lot in most cases
  • core selection: UDP at 10G is typically CPU limited, so its important to pick the right core. This is particularly true on Sandy/Ivy Bridge motherboards.

Do take a look at the article UDP Tuning