ANR Datazero DATAcenter with Zero Emission and RObust management using renewable energy October 1st, 2015 March 31st

ANR Datazero DATAcenter with Zero Emission and RObust management using renewable energy October 1st, 2015 March 31st 2019 Jean-Marc.Pierson@irit.fr 1

An innovative datacenter model» Adapting the IT load to the Power available and» Adapting the Power to the incoming IT load while avoiding unnecessary operations and materials and using a mix of energy sources 2

Target and Outcome Small and middle size datacenters Cloud / Virtualization (1000 m2, 1 MW) Rethink datacenter with alternative power 3

First, we need an electrical and IT connection» Traditionally, a lot of redundancy is present to prevent shortage of power feeds and absorb peak IT demands» How to limit the redundancy while keeping safety?» Use only critical redundant material» How to limit the losses in the electricity conversions and transports?» Connect directly Direct Current Busses with onsite electricity production 4

Classical Electrical Schema: Up to N+1 (Green+Grid) 5

Innovative electrical Schema: Up to 2N Green 6

Second, we need a mix of energy sources» Each source has its own operational parameters, power production efficiency, delay to engage, price,» Which energy sources? How many of each type?» How to operate them at their best?» How to use the power grid?» How to anticipate the power production from sun, wind,?» How do energy storage with batteries and fuel cells provide complementary solutions?» What is the impact of their ageing on the power production elements? 7

We need to profile the Power sources First step : using data measurements (Data Carac), making polynomial approxima;on to extract a model and its parameters, scalable on different sizing Second step, depending on the sizing and reference power to deliver, furnishes discrete equa;ons giving at each discrete step ;me t, what are the current I(t), voltage V(t), state of charge SoC(t) and efficiency/losses of each source. 8

Towards modelling Solar Panels 9

Towards modelling Solar Panels 10

Photovoltaic modelling Different coefficients of third order polynôme (ax3+bx2+cx+d) for each PV technology (5 PV technologies tested) 11

Fuel Cells modelling A set of 3 third order polynoms (ax3+bx2+cx+d) corresponding to 3 specific behaviours. 1 Ac;va;on zone P1 (0A-18A) : 2 Ohmic zone P2 (18A-50A) : 3 Diffusion Zone P3 (50A -85A) FC Stack Voltage (V) 12

Then, we need computers in a datacenter» How do they consume power? What is the link between their usage and their electricity consumption?» What is the impact of the operating system or the hardware tuning?» What is the impact of using Cloud Computing technologies? What is the costs of hypervisors? What is the cost of migration of tasks on the servers? 13

We need to profile the IT Old Intel Xeon Processor E5335 CPU % PSU #1 PSU #2 Intel Xeon Processor E5-2630 OS control power profile Intel Xeon Processor E5-2630 Dynamic power profile 14

Impact of virtualisation and energy savings parameters of servers HP Static low power HP Dynamic power saving 60 70 Power consumption (W) 45 30 15 0 0 23 45 68 90 Power consumption (W) 53 35 18 0 0 23 45 68 90 CPU Activity (%) CPU Activity (%) HP OS control MS Hyper-V HP OS Control Vmware ESXi Power consumption (W) 70 53 35 18 0 0 23 45 68 90 Power consumption (W) 60 45 30 15 0 0 23 45 68 90 CPU Activity (%) CPU Activity (%) 15

Impact of Virtual Machine Migration 16

Then, we need applications on these computers!» Does an application consume power? No, they consume resources (CPU, memory, ), that translates to power consumption.» And applications consume power differently depending on their (obtained or targeted) performances 17

We need to profile the applications An applica;on is modelled by: Time of arrival Resource consump;on profiles over ;me Processor, memory, IO, network Policy related informa;ons (scheduling priority, ) A workload is in SWF format (for each task: arrival time, profile, requirements) 18

Application Profiling: an XML description of the profile <resourceconsumptionprofile> <resourceconsumption> <referencehardware>generic Intel</referenceHardware> <degradationlevel>none</degradationlevel> <duration>pt11s</duration> <behaviour name="userload"><value>0.000000</value> </behaviour> <behaviour name="systemload"><value>0.000000</value> </behaviour> <behaviour name="vsize"><value>11173888.000000</value></ behaviour> <behaviour name="rss"><value>242.000000</value> </behaviour> <behaviour name="minfault"><value>97.000000</value></behaviour> <behaviour name="majfault"><value>0.000000</value></behaviour> <behaviour name="ioread"><value>0.000000</value></behaviour> <behaviour name="iowrite"> <value>0.000000</value> </behaviour> 19

From one application to profiles 20

From one application to profiles Several runs with different parameters > several profiles 20

Finally» We have electrical connections» We have some (renewable) power sources» We have some computers, with their applications» How to make all these elements coexist?» Traditionally: consider all the elements at once, and optimise the global problem!» We believe in an alternative way: the negotiation! 21

We need to negotiate Data center Electrical infrastructure Power Electrical Grid H 2 Cell Optimizations Scheduling Optimizations Electrical management Negotiation IT management Jobs Users 22

Negotiation Module Example 1st approach based on attraction concept: Representing the interest of a system for a proposal - interest for the electrical part to produce 2000W from 11am to 1pm? - interest for the IT part to execute a task at 11am? Abstracted to a float number in [-1, 1] - 1 : absolute reject (a very high penalty) +1 : a very high attraction 0 : neutral intermediate values available Attraction values depend on the possibilities 23

Negotiation: IT attraction Flexibility is given with the deadline 24

But how to assemble and test what we propose?» We can t build a 1 MW datacenter (with the money we have :))» But we can make smaller experiments to validate the models (e.g. power production models, servers consumption, )» AND we integrate results in a platform, mixing both simulated parts and real platforms 26

Middleware, Simulation, Real Hardware Workload Environnemental conditions Power Grid information DCWorms Data Center Simulator (DCWorms) Cloud Open Stack (Grid5000) (EATON prototype testbed) IT Decision module DATAZERO Middleware Platform Negotiation module Power Decision module Power Hardware In the Loop 27 27

As a result, we have a Power Dynamic Data Center with Long Term Green Power Storage Challenge 1. Making demand and envelope constraints coincide Challenge 2. Sizing the infrastructure Challenge 3. Commanding electrical converters Challenge 4. Scheduling IT load Challenge 5. Managing thermal burden Challenge 6. Keeping the complexity at pace Challenge 7. Developing a simulation toolkit 28

Conclusion An ongoing work Still about 3 years to go About 25 researchers involved Keep in touch! Jean-Marc.Pierson@irit.fr 29