A Low-Power Active Gateway for Web Services

A Low-Power Active Gateway for Web Services
Amal Fahad Zhuan Chen Kai Shen Jeffrey Bigham
Department of Computer Science, University of Rochester
Active Gateway
We propose hosting client-bound network services on existing low-power gateway routing device.
Client Proximity
Low-power systems
- As a complementary to clouds, near-client
network services provide unique benefits.
- Energy efficiency: the embedded platform
with low-power components and flash drives.
- Many applications desire such client-proximity
- proxies that offload certain work from clients
- personalized web interfaces
- local group collaboration
- Ease of deployment: small physical size,
co-location with gateway routing.
Internet
Clients
- Thrifty resource provisioning: each host
typically serves a small user population.
The wireless active
gateway running Linux
and applications
Clients
0.008
0.006
0.004
0.002
0
3
2.4
1.8
1.2
0.6
0
1
2
4
8
Number of concurrent requests
1
2
4
8
Number of concurrent requests
Figure 2: Client-perceived average response time.
20
15
10
5
0
1
2
4
8
Number of concurrent requests
Energy per req. (in Joules)
Energy per req. (in Joules)
Low−power gateway
(A) Squid energy usage
Evaluation
Compared with that on conventional machines, applications running on the low-power active gateway
gain large energy efficiency (Figure 3) and suffer from certain performance degradations (Figure 2).
Platforms
- The low-power active gateway
500MHz AMD Geode processor
256MB SDRAM, 16GB CompactFlash
5.9 Watts of power at peak load
Applications
- Squid: a web caching proxy (I/O-intensive).
Conventional server
(B) WebAnywhere energy usage
20
- The conventional machine
2.66GHz Pentium 4 processor
512MB SDRAM, 7.2KRPM ATA disk
73 Watts of power at peak load
15
10
5
0
1
2
4
8
Number of concurrent requests
- WebAnywhere: a centralized web-based
screen reader (computation-intensive).
All on low−power gateway
Figure 3: Per-request energy usage.
Response time (in secs)
(A) Remote server in Seattle, US
To better utilize the low-power active gateway, we need to mitigate the cost of energy efficiency.
- Interference among multiple applications
running concurrently.
- Unfairness read/write performance on the flash
drive.
- Limited resource availability (e.g., small memory)
may lead to more significant competition or even
denial-of-service.
Towards Heterogeneous Platforms
- Load management and request forwarding
decision among different platforms.
- e.g., adaptive offloading for WebAnywhere
between the low-power active gateway and the
remote cloud server, with consideration of both
performance (Figure 4) and energy efficiency
(Figure 5).
3
2
1
0
Adaptive offloading
(B) Remote server in Shanghai, China
1
2
4
8
Number of concurrent requests
3
2
1
0
1
2
4
8
Number of concurrent requests
Figure 4: Client-perceived average response time for
WebAnywhere on local gateway and remote cloud.
All on low−power gateway
All on remote server
(A) Remote server in Seattle, US
Energy per req. (in Joules)
Exploring Quality-of-Service
Challenges for Local Resource Management
All on remote cloud
Response time (in secs)
0.01
(B) WebAnywhere performance
3.6
10
Energy per req. (in Joules)
(A) Squid cache hit performance
0.012
Figure 1: The low-power wireless active gateway
and a prototype on the library bookshelf.
Conventional server
Response time (in secs)
Response time (in secs)
Low−power gateway
8
6
4
2
0
1
2
4
8
Number of concurrent requests
Adaptive offloading
(B) Remote server in Shanghai, China
10
8
6
4
2
0
1
2
4
8
Number of concurrent requests
Figure 5: Per-request energy usage for WebAnywhere
on local gateway and remote cloud.