SeaStar Interconnect: Balanced Bandwidth for Scalable Performance

From AcaWiki
Jump to: navigation, search

Citation: Ron Brightwell, Kevin T. Pedretti, Keith D. Underwood, T. Hudson (2006/07/05) SeaStar Interconnect: Balanced Bandwidth for Scalable Performance. IEEE Micro (RSS)
DOI (original publisher): 10.1109/MM.2006.65
Semantic Scholar (metadata): 10.1109/MM.2006.65
Sci-Hub (fulltext): 10.1109/MM.2006.65
Internet Archive Scholar (search for fulltext): SeaStar Interconnect: Balanced Bandwidth for Scalable Performance
Tagged: Computer Science (RSS) computer architecture (RSS), computer networks (RSS)





  • Design a {balanced, scalable, reliable} interconnect for the Red Storm supercomputer.
    • Balanced: Good bytes-to-flops ratio at each node.
    • Scalable: Scale up to Red Storm supercomputer.
    • Reliability: Transmit correct data.


  • SeaStar
    • Architecture
      • SeaStar figure1.png
      • SeaStar is oversubscribed. Each router can only process 256 concurrent streams, but the system needs to scale up to 1000s of streams at peak.
    • Operating system integration
      • Host-based mode: requests only served from host's kernel. Can trust the messages without verifying, but needs to involve kernel during transmissions.
      • NIC-based mode: requests served from applications. Can't trust messages, but can process them without involving the kernel.
        • Trusted portions of messages have to be kept in kernel memory.
    • Flow protocol
      • Needs to hide transmission failures.
      • Assume the links themselves are always there, so we only need to handle failures stemming from "overprovisioning".
        • In order to detect those failures, routers have to NACK packets they can't respond to. This involves creating yet more traffic on an already congested network.
      • Data and control use different channels. This prevents deadlock; control can go even if the data is stuck.
      • In the future, should use back-off scheme.
  • Programming interfaces: