Exploring the Future Potential of AI-Enabled Smartphone Processors
Green Computing Observatory
1. The Green Computing
Observatory
Cécile Germain-Renaud1, Fredéric Fürst2, Gilles Kassel2, Julien Nauroy1,
Michel Jouvin3, Guillaume Philippon3
1: Laboratoire de Recherche en Informatique, U. Paris Sud, CNRS, INRIA
2: Université Picardie Jules Verne
3: Laboratoire de l’Accélérateur Linéaire, CNRS-IN2P3
2. Outline
— Contexts
— Sensors
— Information Model
— Scientific issues
— Conclusion
The Green Computing Observatory GCG 2011
3. Motivation and Goals
— Energy usages are complex systems
— Sophisticated HW/SW mechanisms eg ACPI, dynamically over-clocking of
active cores, and other optimisations based on on-line statistical monitoring.
— Interaction with local cooling provisioning (eg. fan speed) and global cooling
— Validating generative or predictive models and policies requires
behavioral models based on real data
— The first barrier to improved energy efficiency is the difficulty of
collecting data on the energy usage of individual components, and the
lack of overall data collection.
GCO monitors energy usage at a large computing center, and publishes
them through the Grid Observatory.
— A second barrier is making the collected data usable, consistent and
complete.
GCO adopts an ontological approach in order to rigorously define the
semantics of the data and the context of their production.
The Green Computing Observatory GCG 2011
4. The GRIF-LAL computing room
— 13 racks hosting 1U systems, 4 lower-density racks
(network, storage), resulting in ≈240 machines and
2200+ cores, and 500TB of storage.
— Mainly a Tier 2 in the EGI grid, but also includes
local services and the StratusLab Cloud testbed
— High-throughput, worldwide workload, analysis-
oriented production facility, accessible
approximation of a data center
The Green Computing Observatory GCG 2011
5. Sensors
1GByte/day at 5 minutes sampling period
The Green Computing Observatory GCG 2011
6. IPMI
— IPMI = Intelligent Platform Management Interface,
— Based on a specialized processor card (BMC)
— 1998: IPMI v1.0, 2001: IPMI v1.5, originally by Intel, HP NEC, Dell
,
— 2004: IPMI v2.0 (matured version of IMPI)
— De facto standard implemented by all motherboard vendors
— Fine grain monitoring of individual system parts: temperatures,
fans, voltages, etc. and much more: Recovery Control (power on/
off/reset a server), Logging (System Event Log), Inventory (FRU
information)
— Why? To contribute to a global approach, e.g. cooling inefficiency
leads to increased fan speed which leads to +20% in power
consumption – vs the “hot servers” trend.
— http://www.intel.com/design/servers/ipmi
The Green Computing Observatory GCG 2011
8. IPMI
— IPMI = Intelligent Platform Management Interface
— The exchange protocols are defined and heavily
documented
— But NOT the sensors (nor defined nor documented)
— At LAL, we have DELL and IBM PowerEdge
motherboards
— Very different sensors: e.g. AVGPower (Watts) vs
PSCurrent (Amps)
— Many inactive (NA), may depend on the BIOS version
The Green Computing Observatory GCG 2011
9. Smart PDU
— PGEP PULTI
— 16 outlets
— Each PDU outlet managed separately
— Query protocol : SNMP
— Embedded Web server
— Issue: last systems are Twin2
— 4 systems in 2U, 8-16 processors
— Useful for calibration
— Not all racks will be equipped
The Green Computing Observatory GCG 2011
10. Ganglia
— De facto standard
— Sensors associated with an OS
— CPU load (average number of processes during a given
duration, 1-2-15 minutes) , Memory (buffered, cached,
free, shared, swap, total) and network usage
— Applies to Virtual Machines as well
— Periodic acquisition of system monitoring information (/
proc), transfer and storage protocols. RDD storage
inadequate for the GCO.
The Green Computing Observatory GCG 2011
12. Information model
— There is no standard for
— The output of the physical sensors
— The integration of computational usage and physical
sensors’ output
— There are standards for
— OS information: Ganglia
— Virtual Machine definition: OVF
— Centralized statistics publication: SDMX (Statistical
Data and Metadata Exchange). Successful experience
of porting to a Linked Data model.
The Green Computing Observatory GCG 2011
14. Ontology – measurement
concepts
— Define the semantics of the data (what is measured?) and the
context of their production (how are they acquired and/or
calculated?)
— Observables are individual qualities of endurants (e.g.
temperature of a component) and/or perdurants (e.g. speed
of the rotation of a fan).
— Observations make use of sensors and data acquisition chains
which are physical and non-physical (software) artifacts.
— Observation values are boolean/numeric/scalar qualia.
— Extensions/adaptations of DOLCE, FOOM (Functional
Ontology of Observation and Measurement) and OGC-O&M
(OGC’s Observations and Measurements standard)
— Work in progress
The Green Computing Observatory GCG 2011
15. Publication: XML files
— At 5 minutes sampling period, 1GB/day.
— Scalable querying w/ Xpath
— Integration capability w/ Xinclude
— Easy conversion to analysis-focused formats eg
matlab w/ XSL
The Green Computing Observatory GCG 2011
17. How to
Get an account Download files
www.grid-observatory.org
The Green Computing Observatory GCG 2011
18. Status and Roadmap
— Acquisition of timeseries and metadata for IPMI,
Ganglia, PDU and temperature are in production
— Examples of raw timeseries for IPMI, PDU and
Ganglia released
— Metadata integration and temperature timeseries,
stable XML schema V1 1T 2012
— Global energy consumption 2T 2012
— Ontology-consistent XML schema (V2) 4T 2012
— Also: rack monitoring
The Green Computing Observatory GCG 2011
19. Non-stationarity
The “physical” process is not stationary
— Trends: Rogers’s curve of adoption
— Technology innovations
— Real-world events
— Experimental discoveries
— Slashdotted accesses
NON-STATIONARITY IS A REASONABLE HYPOTHESIS
BUT PRECLUDES NAÏVE STATISTICS
The Green Computing Observatory GCG 2011
20. Intelligibility
How to build knowledge?
— Supervised learning? No
reference, too rare experts
— Let’s build it on-line! Model-
free policies e.g.
Reinforcement Learning!
— Unfortunately, tabula rasa
policies and vanilla ML Exploration/exploitation
methods are too often tradeoff
defeated [Rish & Tesauro
2006).
The Green Computing Observatory GCG 2011
21. Methods
Intelligibility: Uncovering Dealing with non
hidden causes stationarity
— Semantic inference [Y. Kim — Segmentation [T. Elteto et
et al. Characterizing E- al. Towards non stationary
Science File Access Behavior Grid Models, JoGC Dec.
via Latent Dirichlet 2011]
Allocation, UCC 2011]
— Adaptive clustering with
— Collaborative Prediction, changepoint detection [X.
Rank approximation [D. Feng Zhang et al. Toward
et al. Distributed Monitoring Autonomic Grids: Analyzing
with Collaborative the Job Flow with Affinity
Prediction] Streaming. SIGKDD'2009]
The Green Computing Observatory GCG 2011
22. With the support of
— France Grilles – French NGI member
of EGI
— EGI-Inspire (FP7 project supporting
EGI)
— INRIA – Saclay (ADT programme)
— CNRS (PEPS programme)
n
— University Paris Sud (MRM
programme)
The Green Computing Observatory GCG 2011
23. Conclusion: Digital Curation
— Establish long-term repositories of digital assets for current
and future reference
— Continuously monitoring a large computing facility
— Providing digital asset search and retrieval facilities to
scientific communities through a gateway
— Data published through Grid Observatory portal
— Tackling the good data creation and management issues, and
prominently interoperability,
— Formal mainstream ontology, standard-aware
— Adding value to data by generating new sources of
information and knowledge
— Semantic and Machine Learning based inference.
The Green Computing Observatory GCG 2011