Upload
nanoterach
View
23
Download
2
Embed Size (px)
DESCRIPTION
Energy harvesting, energy transfer, nanotechnology
Citation preview
Energy- and Thermal-Aware Design of Many-core Heterogeneous Datacenters
David Atienza, Luca Benini, Edouard Bugnion, Babak Falsafi, Marcel Ledergerber, Fabrice Roudet, Patrick Segu, Lothar Thiele and John Thome
2016 Nano-Tera.ch Annual Meeting
Datacenter Energy Trends Not Sustainable
2
Bill
ion
Kilo
wa
tt h
ou
r/ye
ar
2001 2005 2009 2013 2017 0
40
80
120
160
200
240
280Datacenter Electricity
Demands In the US [Energy Star]
A Modern Datacenter (DC)
Density per rack: 8kW-15kW
[IDC]: Mega DCs is
70% of all DCs in 2018
Change energy increase trend, how to “flatten” it?
Modern datacenters (DC) increase power demands 20 MW
Power is beginning to clearly dominate costs in datacenter management
In Switzerland, 3-4% of all electricity, growing at >20%
Swiss industry is heavily based on services and requires significant IT support
Approach: System-Level Energy-Efficient Integration
3
Multi-scale system engineering to cover complete datacenter re-design Balanced industry-academia participation: 3 large Swiss-based industrial partners
Evaluation of proposed approaches with experimental prototypes
ESL-EPFL
IIS-ETHZ
DCSL-EPFL
PARSA-EPFL
TIK-ETHZ
LTCM-EPFL
WP1: Ultra-Adaptive Server Technologies
4
Analytical models for 1st-ever 4-core in 28nm RVT FDSOI [Solid-State Elect.‘16]
Physically validated energy, performance and temp. confirm energy-scalability of the approach
Evaluated advanced power management to compensate process and temperature variation
28nm FD-SOI
Very effective compensation of leakage and dynamic
power in complete temperature range by VBB or VDD
WP2: Energy-Scalable Server Architectures
Adaptive multi-core design with FDSOI [DATE’16]:
Efficient operation between 200 MHz and 2.5 GHz
Fast change of operating V/F point
Services response time not requiring peak processing
Scale-out loads can operate at 200-500 MHz
VM loads can operate at less than 1 GHz
5
Near-threshold (FDSOI) enables
scalable many-core servers
log(mW)
1GOPS/mW
40nm
log(GOPS)
28nm FD-SOI
Goal: reduce «bending up» of Server’s energy curve
28nm
QoS Limit
WP3: Energy-Free Cooling
6
Built low-height thermosyphon prototype 15-cm prototype proves rack compatibility
Effective use of gravity to remove heat
Mini-thermosyphon demonstrator built [ITHERM’16]:
Validated approach for a 2U server blade,
design done for HP blade (donated by Yahoo)
WP4: Integrated Datacenter Control
Multi-objective controller for geo-distributed data centers Global phase: Virtual Machines (VMs) clustering
Local phase: VMs allocated using CPU-load and data correlation
Green controller: on-line management of energy sources
Significant improvements in cost, energy and perf. Up to 15% in energy and 35% in costs ($$$)
Global phase
(VMs clustering)
Local phase
(VMs
allocation)
Green
controller
DC 1
Local phase
(VMs
allocation)
Green
controller
DC N
Scalable approach tested with more than
12’000 servers distributed in 3 data centers
7
Server and Rack Management Prototype Built with Eaton
Installation at Eaton facilities in Le Lieu Real-time monitoring and management
Reduce peak temp. on multi-rack system
Developed architecture compatible with Eaton’s servers and racks products Evaluations with larger IT setups on the way
8
Remote access possible for testing of
new YINS management strategies
Conclusion
Successful development of prototypes to validate YINS’ research ideas Validated use of 28nm FDSOI tech. to build energy-scalable server architectures
Validated low-height thermosyphon passive cooling system
Developed multi-objective controller for virtualized geo-distributed datacenters
Future work: continue research on inter-WP cross-layer optimizations WP1: adaptive server technology with integrated power and cooling delivery
WP2: optimize memory hierarchy (DRAM, caches, etc) and OS customization
WP3: develop micro-thermosyphon demonstrator in HP blade
WP4: expand experimental validation of multi-rack and datacenter power management
9