The performance of mobile phone processors has been steadily increasing, causing the performance gap between server and mobile processors to narrow with mobile processors sporting superior performance per unit energy. Fueled by the slowing of Moore’s Law, the overall performance of single-chip mobile and server processors have likewise plateaued. These trends and the glut of used and partially broken smartphones which become environmental e-waste motivate creating cloud servers out of decommissioned mobile phones. This work proposes creating a compute dense server built out of used and partially broken smartphones (e.g. screen can be broken). This work evaluates the total cost of ownership (TCO) benefit of using servers based on decommissioned mobile devices and analyzes some of the architectural design trade-offs in creating such servers. You may access the full article (free) here: https://www.usenix.org/conference/hotcloud17/program/presentation/shahrad

1. Towards Deploying Decommissioned Mobile Devices
as Cheap Energy-Efficient Compute Nodes
Mohammad Shahrad and David Wentzlaff
Monday, July 10, 2017

2. Growing Dominance of Smartphones
2
[Gartner, IDC, Zuberbühler Associates AG]
[Ericsson Mobility Report 2016]
2015 2021
3.2
billion
6.3
billion
Smartphones
PCs
Tablets

3. Decommissioned Smartphones
3
Physical Damage
Expanding OSs
New Features
Fashion
Bulk Recycling
Electronic Waste
Refurbish

4. Reusing smartphones as compute nodes?
4
Using the computational capacity of decommissioned devices

5. Outline
5
Why using mobile SoCs now?
How to integrate mobile devices into data center fabric?
What to do with such mobile devices in data centers?
Does such deployment make economic sense?

6. Trend 1: Commodity vs. Mobile Performance
6
[Rajovic et al., Supercomputing with commodity
CPUs: Are mobile SoCs ready for HPC?, SC ’13]

7. Trend 2: Diminishing Performance Growth
7
End of Moore’s Law is starting to kick in
for mobile SoCs.
The performance gap between a new
vs. a 3-year old phone is decreasing.
8,000
80,000
CPUMarkRating
Release Date

8. Extending Effective Lifetime
to Improve Carbon Footprint
8
[Ercan et al., Life cycle assessment of a smartphone, ICT4S ’16]

9. Proposed Integration Fabric
9
Mobile Device Cage
(180 × 80 × 9 mm)
Fans
Router
Power Supply
2U Height
84 × 470~~ SoC TDP = 3.5 W

10. Networking
10
1. USB Tree + Master Node
101010
(a) USB Tree A (b) USB Tree B
1010
Master Node Master Node
Smartphones
3. Intra-server WiFi
Supply
Low SNR and high congestion.
Low
BW High
Energy
High
Latency
2. USB On-The-Go (OTG)
USB OTG Ethernet
Ethernet
Switch
Upstream
Network
Ethernet Switch Cost (price+space)
High BW per Device
No Need for a Master Node
Establish virtual Ethernet through master node
Master becomes single point of failure

11. Mobile Batteries as Distributed UPSs
Batteries could be used to perform
power capping in data centers.
Mobile batteries are designed for
high-density energy storage.
Average mobile battery: > 3100mAh
After 3 years (15% dep/yr): >1900mAh
84 device per server: ~160Ah
(4x-8x the typical UPS density)
11
Time
PowerDraw
Power Cap
Battery Power Draw
Battery Recharge

12. Targeted Applications?
12
Non-CPU Elements of Mobile SoC
Lower Reliability
Lower CPU Performance
High Relative I/O Bandwidth

13. I/O-intensive Application
13
Smaller cores generally have
better I/O to compute ratio.
Mobile cores have better
performance-per-Watt.
Prior examples:
Fast Array of Wimpy Nodes [Andersen et al., SOSP ’09]
• Using low-end embedded CPUs in a key-value store system.
• Two orders of magnitude more queries per Joule.
Web Search Using Mobile Cores [Reddi et al., ISCA ’10]
• Comparing Atom to Xeon processors or perform web search.
• Better energy efficiency
• Worse QoS

14. Low-end VM Provisioning
14
t2.nano
t2.micro
f1-micro
g1-small Multiple tiny burstable instances can be provisioned
on each mobile device.
Each mobile device (on average):
• > 2GB of memory
• > 5 cores
vCPU Mem (GB)
1
1
0.5
1
vCPU Mem (GB)
0.2
0.5
0.6
1.7

15. GPU-accelerated Dwarfs
15
Lower-reliability Nodes
•Current GPU-accelerated IaaS instances use
beefy GPUs (e.g. EC2 P2 instances)
•Mobile SoCs are equipped with smaller GPUs
•Enabling low-end GPU acceleration for small
VMs
•Dynamic reliability platforms use
resource heterogeneity for cost saving
while meeting SLAs
•Mobile devices can be used as low-cost
low-reliability nodes

16. Does it make economic sense?
16

17. TCO Comparison with a Similar-density Server
17
Two simplifying assumptions:
1.Parallelizable workload (using Geekbench 4 cross-platform benchmark)
2.Considered eBay retail price for used phones (could be much cheaper)
TCO: Total Cost of Ownership

18. 6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
Monthl
6 12 18 24 30 36 42 48 54 60
Lifetime (month)
0
1000
2000
3000
4000
5000
6000
MonthlyTCO($)
A, A
=100%
B, B
=100%
TCO A
<TCO B
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
ServerBLi
A
=100%, B
=100%
TCO B
<TCO A
TCO A
<TCO B
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
1000
1200
1400
MonthlyTCO($)
A, A
=45%
B, B
=45%
6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
1000
1200
1400
MonthlyTCO($)
A, A
=45%
B, B
=77.5%
6000
A
=45%, B
=45%
TCOB
<TCOA
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
A
=45%, B
=77.5%
TCO B
<TCO A
TCO A
<TCO B
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
A
=100%, B
=100%
60
TCO comparison with a Similar-density Server
18
6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
1000
1200
1400
MonthlyTCO($)
A, A
=45%
B, B
=45%
6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
1000
1200
1400
MonthlyTCO($)
A, A
=45%
B, B
=77.5%
6000
A
=45%, B
=45%
TCOB
<TCOA
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
A
=45%, B
=77.5%
TCO B
<TCO A
TCO A
<TCO B
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
A
=100%, B
=100%
60
A: Lenovo Flex server
B: Our proposed server
𝛿: annual depreciation rate
(determines the residual value)

19. Thanks for your attention!
19
Next slide: Discussion Topics
101010
(a) USB Tree A (b) USB Tree B
1010
Master Node Master Node
Smartphones
8,000
80,000
CPUMarkRating
Release Date 6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
1000
1200
1400
MonthlyTCO($)
A, A
=45%
B, B
=45%
6 12 18 24 30 36 42 48 54 60
Lifetime (month)
400
600
800
1000
1200
1400
MonthlyTCO($)
A, A
=45%
B, B
=77.5%
3000
4000
5000
6000
nthlyTCO($)
A, A
=100%
B, B
=100%
A
=45%, B
=45%
TCOB
<TCOA
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
A
=45%, B
=77.5%
TCO B
<TCO A
TCO A
<TCO B
6 12 18 24 30 36 42 48 54 60
Server A Lifetime (month)
6
12
18
24
30
36
42
48
54
60
ServerBLifetime(month)
A
=100%, B
=100%
TCO B
<TCO A
30
36
42
48
54
60
Lifetime(month)

20. Some Discussion Topics
20
•Reliability of used mobile devices under data center workload
•The best management scheme (centralized, distributed, server-level, etc.)
•Using SoC accelerators for domain-specific applications. (security, neural net, etc.)
•Dealing with extreme heterogeneity
•Security concerns with used phones