PERI-DB Meetings 2008

From PERI

Contents 1 Introduction 2 Notes 2.1 PERI-DB 9/17/08 2.1.1 Data collection 2.1.2 Action items 2.2 PERI-DB 9/3/08 2.2.1 Data collection 2.2.2 PERI-XML 2.2.3 Action items 2.3 PERI-PI and PERI-Architecture 8/27/08 2.4 Teleconference 8/13/08 2.4.1 Action items 2.5 Teleconference 7/30/08 2.6 Teleconference 5/21/08 2.7 Teleconference 5/7/08 2.8 Teleconference 4/23/08 2.9 Teleconference 4/9/08 2.10 Teleconference 4/2/08

Introduction

Year of 2008 meeting notes from the PERI-DB group.

Notes

PERI-DB 9/17/08

Agenda: Go over action items for last call

Attending: Rob Fowler, Kevin Huck, Wyatt Spear, Allen Malony, Xingfu Wu, Dan Gunter

Data collection

Infrastructure: perfdb.renci.org is now a virtual host, Dan is still in the process of migrating the database from the "old" perfdb.renci.org (which is now called perfdb2.renci.org, btw). Will send out mail / update wiki when done.

Hardware counters: Rob suggested that consistent semantics are really only available for a few counters -- instructions, cycles graduated, bus cycles, and maybe last level cache misses. Suggestion was to bring this up again on the architecture/peri-pi call later this morning.

Action items

• Continue asking for feedback from Boyana on requirements
• Finish migration to new perfdb.renci.org database; send out instructions, etc. when done
• Attend PERI-PI call and cc: PERI-DB list with relevant info

PERI-DB 9/3/08

Agenda: Discuss data collection, maybe PERI-XML

Attending: Kevin Huck, Shirley Moore, Bernd Mohr, David Skinner, Dan Gunter

Data collection

• • Shirley has run GTC on Jaguar with timings and with TAU from 32 .. 256 (pow2) processors
  • Which HW ctrs to collect?
    • minimal set (across all): FLOPS, L1 -- should agree on this
    • other: Tot Ins, TLB misses
  • TAU results, PPK files on Shirley's computer

PERI-XML

Nobody has done much since the last meeting. Bernd raised the point that this work is less interesting to tool developers if it is narrowly focused on providing PERI autotuning with what it needs; he is in "wait and see" mode, will talk with Martin (Schulz) in 2 weeks. Other possible venues: grid forum, "Point" initiative in Europe.

Action items

(Dan will..)

• Send email to PERI-Arch group asking for minimal set of HW counters
• Give Shirley an account on perfdb, post uploading instructions
• Post details about how to access data from TAU portal (through LBL tunnel)
• Nag Boyana for some autotuning's requirements for profiles
• Talk to Rob Fowler about opening a hole in perfdb firewall to avoid the hop through LBL

PERI-PI and PERI-Architecture 8/27/08

Attending: PI's (regrets: Bronis)

Notes related to data collection:

• Sameer: TAU data on S3D on Jaguar, 4000 cores
• FLASH on BG/P (Boyana). New code release mid-Sep. at that point: stable. But they don't consider the current version "stable". Can release to PERI immediately. Will email and post on wiki with instructions/status/plans. Two scenarios (input decks) to be used.
  • FLASH tutorial at AHM in Sept.? Bob: 1 hour certainly OK, not clear how much time needed; if need 4, carve out as parallel session.
  • (Bob) Timeline: working backwards from Sept. 15, to get something to show to Fred, Paul should construct plan of what to run.
  • Sameer volunteered to help with runs on Jaguar.
• S3D Heike is collecting data on Jaguar (not yet on Wiki). Email exchange with Jackie: new version is not ready for BG/P. Status of BG/P port not clear. So for BG/P and Jaguar at same time only old version of code is available, which is what Jackie says to do. Heike still waiting for BG/P account, paperwork filed late last week.
• GTC (Shirley). Trying to get working version of code. Shirley talking with Scott Klasky, who apparently has the "Yong" GTC version in a directory, but referred her to Chen. Shirley has been going in circles since SciDAC meeting, but hope to resolve soon. As soon as she gets working code, she can do runs. Developers claim they will have an operational version by the end of the week.
  • (Bob) What about John M-C's results? answer: he has a different version. New ("Yong") version has the desired science.
  • Shirley needs to finish getting an Intrepid account.
• To summarize: two codes, one on each machine, running. All the others are "just about ready" to be run.

Teleconference 8/13/08

Attending: Dan Gunter, Kevin Huck, Martin Schulz, David Skinner, Xingfu Wu

Agenda: Status of Data Collection for PERI-Arch

• Databases
• PerfTrack collection how-to

Status of PERIXML

Discussions:

Topic #1: Data collection

Most active participants in this were not on the call, so Dan gave a brief update.

• Data is loaded into MySQL database for PerfTrack and TAU

Topic #2: PERIXML:

Worked through Dan's IPM example sent out before the call. Following are the major points of the discussion.

• pid doesn't belong there -- should be job id
• pid, code name: duplicates OK
• host to rank mapping, not here
  • as in TAU, need perhost and per-run metadata: this could be done in "space" dimension
• need way to map different dimensions to each other
• derived vs. "raw" metrics. need an algebra to describe this?
  • no, just consistent notation: initially being read/processed by humans
• Line between "code" and "state" for functions?
  • "state" is anything you can't determine at compile time? Need clearer rules for how to split it up.
  • Rule = everything goes under code, unless the information needed is run-time. In which case it goes both places.
• Each dimension can only be represented once in a given tuple. At most the length of each tuple is four (if we have 4 dimensions).
  • Hierarchy of dimensions in 'DATA': there are multiple ways to nest things. Nesting and/or include a tuple for the data point.
  • Hierarchy ends up being a compression method, not strictly required
• Hierarchy of dimensions in 'DIMENSION': If the dimension is a non-leaf, the aggregation operation needs to be specified.
  • Need metadata that maps these hierarchies to each other.

Action items

• Dan will give Kevin an account on the host (perfdb.renci.org)
• Dan will re-work IPM example based on discussion and send out
• Martin will try to generate an O|SS example, although he will be on travel for the next month.

Next teleconference in two weeks, same time and number.

Teleconference 7/30/08

Attending: Dan Gunter, Rob Fowler, Shirley Moore, Bernd Mohr, Boyana Moore, Martin Schulz, Kevin Huck, Xingfu Wu, Karen Karavanic, David Skinner, Alan Malony

Agenda:

• Data collection plans for PERI "Architecture" tiger team. From Patrick Worley, 7/25/08:

Dan Gunter, Shirley Brown, and Rob Fowler should set up and manage
a performance data repository. In particular, they should help the
code groups to upload the data, hopefully training them in the
process. As part of this, the system, application, benchmark instance,
and performance tool/data-specific metadata needs to be defined and
tagged to the data.

• PERIXML schema. See Martin Schulz's PPT slides

Discussions:

Topic #1 (Data Collection):

• Rob: Have blade, sysadmin provisioning it today. Dan will get sudo on it.
• Karen: Ready to install PerfTrack
• Shirley: Code groups -- more like PERI people. For GTC, we're already collecting some.
• Karen: Have client GUI that can look at database from people's local machines

Topic #2 (PERIXML Schema):

• Rob: Compactness is important. Need to deal with inlined code, which is common. Need more formal properties of this representation (projections of sample space) for composition of contexts -- will write this down and send around.
• Martin S.: From slide #7, "regions" instead?
• General issue seems to be a linking of hierarchies..
• Rob: Need set of worked out examples
• Karen: need to define goals first. What tools / kinds of data are being included?
• Martin S.: Any profile data.
• Rob: Use-case for interchange?
• TAU, Xingfu->Prophesy, Bernd->CUBE viewer
• Rob: Autotuning tools should be main "customers" ... though existing tools will intermediate
• No "requirements" list yet for this
• Boyana: (from talks with TAU) there is some

Action Items:

Topic #1 (Data Collection):

• ((Karen)) Put (brief) instructions for PerfTrack on architecture wiki page
• ((Dan/Shirley)) Cross-reference data-collection, instructions, GTC/etc. pages

Topic #2 (PERIXML):

• Wiki page with list of targeted tools and machines
• Subgroup to come up with list of examples and send around to larger group
• Boyana will write down requirements use-case for their work (Dan will poll for this, then post to wiki)

Teleconference 5/21/08

Attending: Kevin Huck, Martin Schulz, Dan Gunter

Agenda:

• schema / paper

Discussion:

• Kevin and Martin plan on meeting F2F in July with John M-C so three of the tools at least (TAU, OSS, HPCT.K.) could hash out what works and what doesn't in the current data schema.
• Dan noted that the latest data schema is in subversion; he will continue working on IPM translation

Next telecon in 2 weeks

Teleconference 5/7/08

Attending: Rob Fowler, Martin Schulz, Xingfu Wu, Karen Karavanic, David Skinner, Dan Gunter

Agenda:

• Paper status
• schema status
• RENCI system

Discussion:

• Paper status: We discussed possible directions and venues
  • using PERI architecture tiger-team data seems distant, as they are months away
  • Journal possibility: detailed discussion of database designs
  • Focus on the format itself, talking about benefits of interoperability
    • General feeling was that this was a better near-term target
• schema status
  • David Skinner (IPM) mentioned he's looking forward to being able to convert to it
  • Still need to finish the performance-data part
• RENCI system (Rob Fowler)
  • Dell blade center, part of which is for research, and the rest of which is to consolidate the existing services
  • from a week or so after now until July: bare metal
  • after July: large RDBMS server with MySQL and Postgres
  • by August: blades with virtual servers (not Xen; commercial product)
  • As long as not too much sysadmin required, we can store data on this! A few people will get logins.

To-do:

• Focus on performance-data part of schema in next few weeks
  • Dan will send out an updated version soon, and we will discuss next time
• Everyone to look for conferences, send links

Next meeting: 2 weeks, normal time.

Teleconference 4/23/08

Attending: Rob Fowler, Kevin Huck, Xingfu Wu, Karen Karavanic, Dan Gunter

Agenda:

• Paper status
• Interaction with PERI architecture Tiger Team
• schema status

Discussion:

• Note that for SciDAC08 conference, deadline is probably about July'08
• For an actual conference paper:
  • Need some results first
  • Perhaps a June target
  • Suggestions welcome
• Interaction with PERI architecture Tiger Team:
  • Apps: GTC, S3D, Flash
  • Being led by Bronis, Pat Worley, ..
  • What should PERI-DB be doing?
    • focus on needs of this team, not general problem
    • "impose discipline on logging of experiments" - Rob F.
    • RENCI can host the DB; has svn server (https), Apache, MySQL on a Linux box they can use
    • Potentially use PerfTrack to collect provenance

To-do:

• Rob to get back with details about RENCI hosting, when available
• Karen will send a tarball of PerfTrack code
• Dan will try out PerfTrack on ANL BG/P and ORNL XT4
  • look at what's necessary to conform to current schema
• Next telecon in 2 weeks, as usual

Teleconference 4/9/08

Attending: Rob Fowler, Martin Schulz, Xingfu Wu, Dan Gunter

Agenda:

• Discuss current paper status

Discussions:

• Paper TO-DO list (assignee in parens)
  • New title: "PERI XML: Standardizing Efficient Information Interchnge Between Performance Tools"
  • Break author list into 2 rows: main contributors, everyone else (Dan)
  • Add references from last year (Dan)
  • Re-re-add section 4.1 (Martin to send text to Dan)
  • Add O|SS ref (Martin)
  • Section 4.3, QCD (Rob F.)
• Longer-term plans
  • Need to pick a conference for a fuller paper with results
• Scheduling
  • Move meetings back to bi-weekly, so next meeting is in 2 weeks
  • Also move meeting time back 15 min. to 2:00 PT (5:00 ET)

Next meeting: Wednesday, 4/23/08, 2pm PST. Note new meeting time

Teleconference 4/2/08

Attending: Dan Gunter, Kevin Huck, Xingfu Wu, Ying Zhang (Martin Schulz was at SciDAC PI meeting in Denver)

Agenda:

• Discuss current document, PERI Database Infrastructure Serving the SciDAC Community

Discussions:

• Need to ask Rob and/or PERI PI's about when/how we put this in as a SciDAC poster
• Need a simplified "Boyana figure"; Dan will do it
• Ying will try and contact Karen to see if she wants to participate in this paper
• Ying will add references at the bottom from last year's paper and everyone will edit
• Finally: this was Ying's last teleconference as erstwhile leader of the PERI-DB group. We'll miss her. Dan Gunter, as the only person left in the group actually funded by PERI(!), will take over.

Next meeting: next Wednesday, 4/9/08