Highest%20performance%20parallel%20storage%20for%20HPC%20environments

1
Highest performance parallel storage for HPC
environments
Garth Gibson CTO Founder IDC HPC User Forum,
I/O and Storage PanelApril 21, 2009
2
IDC Storage Panel Question 1

• With the formalization of Parallel NFS as a
  standard, what steps are being provided to enable
  this to be hosted on current (and future)
  platform choices?
• Panasas has contributed two engineers to the open
  source Linux code effortfor pNFS. Code has
  started to be included in Linux (2.6.30) and
  RedHat is beginning to grease the path into
  Fedora and RHEL.
• Panasas has an implementation working with the
  current Linux head of line code.
• Los Alamos National Lab has started testing with
  Panasas on a 128 node cluster. Ultimate testing
  may be done on the worlds fastest supercomputer
  today, LANLs Roadrunner, which runs PanFS now.
• LANL also has plans to test with IBM, BlueArc and
  maybe with LSI.

3
IDC Storage Panel Question 2

• What are the tools available to help optimize
  pNFS from the application level down?
• Panasas collaborates with the DoE Petascale Data
  Storage Institute (PDSI) on its agenda to clear
  the path to Exascale. PDSI has built tracing
  tools and benchmarks for complex HPC codes that
  will be applied to pNFS.
• In the toolkit of NFS development the most
  important tool is frequent interoperability
  testing at Connectathon face-to-face engineering
  meetings. For debugging, Wireshark has added
  NFSv4.1 (pNFS) trace specific parsing rules.

4
IDC Storage Panel Question 3

• We are all facing complexity and cost issues.
  With IB or 10 GE (40/100 GE), where should the
  HPC community focus its resources?
• Panasas supports 1GE, 10GE, and IB connectivity.
  We believe 10GE is the most compelling long term,
  especially with data center ethernet protocol
  enhancements, but IB is quite effective today.

5
IDC Storage Panel Question 4

• There are too many standards, interconnects,
  media layers today. iSCSI/FCoIB/FCoE/FCoCEE have
  all been touted as the solution. Is this even
  relevant in the HPC arena? Is fragmentation the
  only choice?
• Panasas believes in converged networks based on
  Ethernet. We use iSCSI today in solutions of up
  to 30 GB/s.
• Converged and Enhanced Ethernet, or Data Center
  Ethernet, is an interesting optimization, but is
  not required.
• Panasas does not see FC or any of its variants to
  be important in HPC. It might be incidentally
  present inside storage boxes, although SAS should
  be fine for this too.

6
IDC Storage Panel Question 5

• What are the top 3 main (technical or human)
  issues in HPC I/O today?
• While high performance parallel file systems now
  support concurrent writing into one file system
  from 1000s of nodes at the same time, large
  numbers of codes are still written for the era in
  which storage was connected to only one node.
  Pushing all I/O through one node does not scale,
  but adapting and optimizing stable code bases
  takes many years.
• Enterprise-class reliability and availability has
  long been absent in high performance, scalable
  HPC systems. Too many solutions are too
  do-it-yourself to be reliable, well integrated,
  well supported AND scalable.
• Multiple divergent business models confuse things
  is scalable storage software developed and
  tested by hackers in their basements, is it a
  loss leader cost center buried in the price of
  compute cluster hardware, or is it a valuable
  asset in itself?