Glenn's Digital Garden

ParaStor F9000

ParaStor is a parallel file system developed by Sugon. A few factoids were disclosed about the all-flash configuration in its IO500 submission:

  • Data is erasure coded at 14+2, and metadata is triplicated across servers1
  • It uses a custom kernel client rather than FUSE or NFS with a distributed lock manager. The server side uses FUSE somehow.2
  • The server software localizes CPU cores, NUMA domains, NICs, and SSDs to minimize latency and variability.3 This sounds very WEKA-like.
  • QOS is supported for both bandwidth and IOPS and can be specified as mins and maxes.2
  • Clients can use node-local SSD and RAM as a “near-computing power cache” with prefetch.2 This sounds very WEKA-like.
  • P2P DMA via “XDS,” which is Sugon’s equivalent to NVIDIA GPUDirect Storage.2
  • It supports both HDD and SSD tiers, with “intelligent data tiering” (whatever that means).2

ParaStor F9000 appliances are dual-controller 2U24, and each controller has:2

Sugon had a few different rack configurations on display at ISC26: a 42U (HDD), 36U (SSD), and 26U.

Footnotes

  1. IO500 - Submissions ↩

  2. For the first time in a decade, China has won two titles: this time it’s not just about performance. ↩ ↩2 ↩3 ↩4 ↩5 ↩6

  3. According to the chief architect of distributed storage at Sugon,2 “For every IO issued by a user, from the moment it is issued, we know which network card it will go through, which memory it will use, which core it will pass through, and which disk it will finally land on.” ↩

Graph View