Notes for Hoffmann cluster researchers

Recap:

During this recent test period in November, researchers confirmed that the new system's hardware and base configuration fully enables the proper functioning of all the software they currently depend on within Helios. We discovered important things that we could not have learned without this testing investment, which in turn will contribute to a more robust setup for your researchers.

Upcoming to do's for Hoffmann researchers, as of 11/27/13:

(1) As soon as possible, ensure researchers' user data on Sol can be completely deleted.

• We will be deleting all the home directories. (They were originally copied there as a testing convenience during our test phase.)
• We do this to enable a clean import of researchers' current production home directories from Helios.
• Lulu has advised researchers of this from before the Thanksgiving break, so hopefully no surprises.

Bottom line: If there are any files on Sol which researchers cannot afford to lose (hopefully not- it was in test mode, and it's not being backed up!), researchers must move that data to Helios right away (deadline?).

2) On a predetermined date, researchers will stop having access to Sol until it is again ready for testing.

• Sol will need to be turned off and worked on by us. (Approximate duration?) Therefore, it will be off-line to researcher.
• Ad hoc permission to select researches (Prasad?) will be granted so they may install and test their software.
• Tasks for us to do:
  • Get the server's versioning hard disk recognized within the new head node.
  • Lulu to install the Intel compiler in the /software/ directory, with the proper licensing.
  • Your group (Prasad?) to install the Materials Studio software in the /data/ Hoffmann/bin/ directory. If Prasad is not available to do this, let's discuss our options- thank you.
    

3) On a predetermined date, researchers will be asked to test select software on Sol.

• Changes to where the Intel compiler and Materials Studio reside (to facilitated debugging and support) must be confirmed to work for researchers.
• Again we are testing on Sol. Thus, the only production work should still be on Helios, not Sol.

REMINDER: Researchers must again ensure their user data on Sol can be completely deleted.

4) On a predetermined date, researchers will again stop having access to Sol until it is ready for production.

•  During this time we will sync researchers' home directories from Helios
  QUESTION for Lulu and Michael: Is this really better than giving each researcher a bare home directory and asking them to move only the data from Helios which they need?

5) On a predetermined date, researchers must plan for 1-2 days of downtime for Helios, after which Sol will be available and in full production.

• We obviously must coordinate this cut-over to ensure no loss of researcher's production data and to minimize the inconvenience of the downtime.
• With Sol in production, we can:
  • add the older compute nodes from Helios, making Sol more and more powerful.
  • once Sol is confirmed "good", Helios's head node can be added as a compute node to Sol.

Thank you!  -ChemIT

Older notes:

Next steps

• Meet to review all options and confirm desired direction and expected timing.
• Review resources. Huayun Gen has cluster management experience, including set-up.

Draft idea

• Create a stand-along cluster using new hardware ($25K for minimum of 3 years operations (to confirm!). Thus, ~$8K/yr in hardware)).
  • Uses new OS and related cluster management software.
  • Install and configure necessary applications.
  • Enable NetID-based access, if possible (limit 2-3 days for a "go/no-go" decision on this functionality)
• Confirm old nodes can successfully be added to that new cluster.
• Migrate users and data to new cluster.
• Migrate old nodes to new cluster.

Unknowns

• Time for install of all necessary applications, many of which are new to Lulu. Then configure, verify, and de-bug new-installation-related issues.
• Whether NetID-based access will succeed. But note that this is not a do-or-die step, thus we will limit the duration of our investigation, with the hope that we can make this happen.

Tasks and estimated timing

• Lulu becomes available ~mid-September or early Oct, as of 8/21/13.
• See unknowns, above, which related to tasks which will obviously take additional time to accomplish.

Other provisioning models and related ideas

• We can walk through rates and scenarios, as appropriate.
• We can meet with CAC since they may be willing to do more with a commitment of $25K than is published with their $400 min. offering.
  • Brainstorming idea: Would they be willing to add hardware to CAC's RedCloud to get a buyer of that hardware a better cost and/or privileged access?

Buy cycles, on demand

Good for irregular high-performance demands, especially if have high peaks of need and long-lasting jobs.

• Buy cycles from CAC (RedCloud, minimum of $400 for 8585 core*hour
  • http://www.cac.cornell.edu/RedCloud/start.aspx
  • 12 cores available at any one time on one system.
    • Can access more than one system at a time, but systems are not linked.
  • $400 (minimum) buys you 8585 core*hours
    • This comes out to ~1 core for an entire year, non-stop.
  • For 96 cores, that's $38.4K for 1 year, non-stop.(They have a max of 96 cores <http://www.it.cornell.edu/about/projects/virtual/test.cfm>.)
    • 96 = 8 nodes, each with dual 6-core procs => 8 * 12 = 96
  • Or, for $25K, that's ~536,562 cores*hours.
    • $25K = $400*62.5 units. And each unit is 8585 core*hours, so 62.5 of them gets you 536,562.5 cores*hours.
    • That comes to ~178,854 core*hours/ yr for 3 years. Which is 20.8 core system running non-stop each year. (Compare to one hardware node, which has 12 cores.)

• CNF, w/ Derek Stuart.
  • A very reliable cluster, per Roald.

• Determine costs, processes, and trade-offs if use another cloud service, such as:
  • Amazon. Amazon AC3?
  • Google. Google Compute?
  • Microsoft. Microsoft Azure?

Host hardware at CAC rather than with ChemIT

Hosting costs at CAC is for basic: Expert initial configuration, then keep the system current, and keep the lights running. Other service charged hourly.

• http://www.cac.cornell.edu/services/hpcsystems.aspx
• http://www.cac.cornell.edu/services/rates/

Per the above rate calculator, the rate for 9 nodes (1 head node + 8 compute nodes) would be $8,291/yr. Or, $24,873 for 3 years for this service.

At current ChemIT rates, 9 nodes would be $321.84/yr. Or, $965.52 for 3 years of service.

• ChemIT rates are set by the CCB Computing Cmt and may change at any time. The rate for a group's single system (in a cluster or not) is $2.98/month, or $35.76/yr.

Table, related to our options