June 1 2018 Notes

Agenda

1. Update on Power API outreach – OpenCompute group (http://www.opencompute.org/wiki/HPC​)
2. Update on new plugins/support from HPE
3. Power API news from HPPAC workshop
4. Voting – Ticket #2 (front matter) and Ticket #3 (PWR_ObjGetSizeName) (https://github.com/pwrapi/powerapi_spec/pulls)
5. Discussion of Roles/Security policy – How do we want to express permissions? Integrate into Spec or demonstrate in community implementation?
6. Any other business

Attendees

Bill Lucnik
Jeff Hanson, HPE
Steve Martin, Cray
Ryan Grant, SNL
Lee Ward
Barry Rountree, LLNL
Matt Kappel, Cray
Ram Nagappan, Intel
Jim Laros, SNL
Kevin Pedretti, SNL
Sid Jana, Intel
Martha for Todd, IBM

Discussion

Open compute group trying to develop open platforms. Invited PowerAPI and GeoPM people. Looking for common api for power management. They seemed more interested in Power API. Ryan is going to lean on them to join. Some vendors were interested in making plugins. Ram asked in they were going to do Redfish and PowerAPI. Ryan said they are trying to figure it out. Since we already have Redfish they can just use that. Ryan to provide links and other details.

HPE provided updated code. There is one out of date package that will need to be updated to a newer provider of that code.

HPPAC Workshop details many people attended last month. Talked to developers on Bull system which has the capacity for KHz sampling on individual nodes. Barry asked what is using that currently. Ryan said there is an auto tuner (readX?) and a monitor (dashboard). [Ryan to put in more details from his notes].

Matt got clearance to submit a pull request (make more code available).

Voting

These are topics we voted on last time and Ryan did what we agreed to.

Ticket number 2 Ryan shared the format with names on front, organizations on the second page

• Cray - votes yes
• HPE - votes yes
• LLNL - votes yes
• IBM - yes
• Intel - yes
• SNL - yes

Ticket number 3 Ryan gave proposal for new function. Ryan asked Steve Martin if this was okay with him as Steve had the most issue last time. Steve said it was spot on with the rest of the API. Kevin quibbled that the order of the name looks odd. Ryan said yes this would set a precedence was this is the first of these functions. Steve like Kevin's change of order. Ryan wishes us to vote but with adding of in three places.

• Cray - votes yes
• HPE - votes yes
• LLNL - votes yes
• IBM - yes
• Intel - yes
• SNL - yes

Barry asked how hard was it to edit the document. Ryan said it went quite well.

Matt's proposal for real time statistics. Basic idea is during implementation/use of RTS you might have std dev on power. [I missed a couple sentences]. Steve's summary is you have to tell it what statistic you want when you create the object. But you might not know ahead what you want. So if we change to create an object and then ask question of the object. Minimizes the number of objects that track the same data. Worst case would be getting the wrong data. Max being less than min with sine wave (for example). Plan is Matt will have a pull request before next meeting. Ryan saw it before and there were not technical issues.

Ryan and new issues in spec. He took the Roles/Security policy issue. How shall we specify. If one had a User/Application view one would get just the job that currently belongs to the user. How fine grained if we are going to specify. Example many HPC is you are the user on a node we could give the whole node. If you got just a subset of the node, how to provide that. Barry thinks it would be more persuasive if it was implemented first and then fed it into the spec. LLNL has the controls it needs so it probably doesn't need to be in the spec. Steve said in Trinity project they did whitelisting and there is policy in implementations. He'd like feedback on how it went. He agrees with Barry in principle. Jeff said we might need a bigger sample space on implementations. Steven wondered what does the spec actually says. Too open ended is bad because it becomes less portable. Doing subnode is hard. Jim said API has some details because they didn't want to push SNL and/or HPC view too much. Ryan said there is value to review implementations and see what is missing. A simple document might result. Ryan asked for a summary. Ryan and Barry will work on a review. Steve/Matt agree to answer questions as needed.