Poster abstract: Enabling reliable and high-fidelity data center sensing.
ABSTRACT RACNet is a sensor network that monitors a data center's environmental conditions at high temporal and spatial resolutions. The data RACNet collects can improve the energy efficiency of data centers, currently one of the fastest growing energy consumers in the U.S. RACNet overcomes the challenge of reliable and low-latency data gathering from dense networks deployed in harsh RF environments through rDCP, a novel network protocol that decouples data collection from topology control. Furthermore, rDCP automatically partitions the network to multiple routing trees, each operating at a different frequency channel. The combination of these features enables rDCP to scale up while maintaining high data yields. Preliminary results from a production deployment of 694 sensors (including 174 wireless nodes) show that rDCP achieves a data yield of 99%, while delivering 90% of the measurements in less than 30 seconds.
- SourceAvailable from: Rodrigo L C Fonseca[show abstract] [hide abstract]
ABSTRACT: Despite being a core networking primitive, collection pro-tocols today often suffer from poor reliability (e.g., 70%) in practice, and heavily used protocols have never been eval-uated in terms of communication efficiency. Using detailed experimental studies, we describe three challenges that cause existing collection protocols to have poor reliability and waste energy: inaccuracies in link estimation, link dynam-ics, and transient loops. In this paper we present CTP, a robust, efficient, and hardware-independent collection protocol. CTP uses three novel techniques to address these challenges. CTP's link es-timator addresses the inaccuracies in link estimation by us-ing feedback from both the data and control planes, using information from multiple layers through narrow, platform-independent interfaces. Second, CTP addresses link dynam-ics by using the Trickle algorithm for control traffic, send-ing few beacons in stable topologies yet quickly adapting to changes. Finally, CTP addresses transient loops by using data traffic as active topology probes, quickly discovering and fixing routing failures. CTP runs on six different mote platforms and we have tested it on four testbeds. In most experiments, CTP achieves 99% reliability, and in some cases 99.9%. In the most chal-lenging testbed, the state-of-the-art collection protocol to-day (MultiHopLQI) achieves 70% reliability: CTP achieves 97%. CTP achieves this ten-fold reduction in dropped pack-ets with 25% fewer transmissions. CTP works seamlessly on top of existing low-power MAC layers. Together, these re-sults suggest that CTP can be the robust, efficient collection layer that so many sensor network applications and protocols need.
Poster Abstract: Enabling Reliable and High-Fidelity Data
Chieh-Jan Mike Liang†, Jie Liu‡, Liqian Luo?, Andreas Terzis†
†Johns Hopkins University
RACNet is a sensor network that monitors a data cen-
ter’s environmental conditions at high temporal and
spatial resolutions. The data RACNet collects can im-
prove the energy efficiency of data centers, currently
one of the fastest growing energy consumers in the U.S.
RACNet overcomes the challenge of reliable and low-
latency data gathering from dense networks deployed in
harsh RF environments through rDCP, a novel network
protocol that decouples data collection from topology
control. Furthermore, rDCP automatically partitions
the network to multiple routing trees, each operating at
a different frequency channel. The combination of these
features enables rDCP to scale up while maintaining
high data yields. Preliminary results from a produc-
tion deployment of 694 sensors (including 174 wireless
nodes) show that rDCP achieves a data yield of 99%,
while delivering 90% of the measurements in less than
Data center energy consumption has attracted global
attention due to the fast growth of the IT industry
and increasing concerns about carbon footprints and
global climate change. Approximately 40% to 70% of
the total energy that a typical data center uses is lost in
the power distribution system or used by environmental
control systems such as Computer Room Air Condition-
ing (CRAC) units, water chillers, and (de)humidifiers [1,
6]. A root cause for this low energy efficiency is the lack
of visibility in the data center’s operating conditions.
For example, when servers issue thermal alarms, data
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
Copyright 2009 ACM 978-1-60558-371-6/09/04 ...$5.00.
center operators tend to decrease the CRAC’s tempera-
ture settings, which might not alleviate the problem .
Wireless sensor network (WSN) technology offers
many advantages for monitoring and control tasks. It is
low-cost, non-intrusive, and can be easily re-purposed.
Furthermore, WSNs require no additional network and
facility infrastructure, simplifying their deployment in
the already complex data center IT environment. At the
same time, data center monitoring introduces new chal-
lenges for WSNs. Temperature can vary by as much as
5◦C over a couple of meters inside a data center, which
means that thousands of sensors might be necessary to
cover a single colocation facility. Moreover, Liang et al.
showed that 83% of the motes in a colocation are within
one-hop distance from each other . The combination
of large scale and high density complicate the task of re-
liable and low-latency data collection. To make things
worse, the RF environment inside a data center is harsh
as it includes multiple metallic obstacles and interfer-
ence from WiFi networks.
RACNet uses custom-made motes, called Genomotes,
that form sensor chains using off-the-shelf USB cables.
These cables deliver measurements, and they also dis-
tribute power from a server’s USB port to all the sensors
on the same server rack. Each chain consists of multiple
sensing slaves and a master that is responsible for relay-
ing the measurements over its radio. An added benefit
of the chain configuration is that it reduces the num-
ber of nodes in the wireless network thereby reducing
contention. Both mote types use an MSP430 microcon-
troller, while the master also has a TI CC2420 802.15.4
radio, 1MB of flash for data buffering, and a battery for
continuous operation in case of server down time.
Figure 1 provides an architectural overview of RAC-
Net’s reliable Data Collection Protocol (rDCP). First,
rDCP uses a distributed protocol for maintaining bi-
directional routing trees (BiTrees) used for data col-
lection. Multiple BiTrees can coexist, each rooted at
a gateway that occupies a different frequency channel.
Membership in individual trees dynamically adapts to
the number of gateways in the network as well as link
Data Download Layer (DDL)
Topology Control Layer (TCL)
Figure 1: Reliable Data Collection Protocol (rDCP) ar-
chitecture.The Topology Control Layer (TCL) main-
tains the tree topology while the Data Download Layer
(DDL) reliably downloads data from individual master
qualities on each channel. Unlike centralized data collec-
tion protocols, such as Koala , the distributed topol-
ogy control nature of rDCP allows it to promptly adapt
to link quality changes and scale to large networks. Sec-
ond, data downloads in rDCP are centrally controlled
by the network’s gateways. This is different from dis-
tributed data collection protocols, such as CTP , in
that nodes wait for gateway requests before streaming
measurements. This centralized coordination minimizes
interference among network flows. At the same time,
the pull-based data download approach allows the gate-
ways to issue commands and negative acknowledgments
used to implement reliability.
To minimize data latency, rDCP attempts to balance
the load among all gateways, defined as the sum of hops
necessary to reach all nodes in the tree. Once a gate-
way’s load exceeds the network average by a predefined
threshold, it calculates two probabilities according to
its load relative to that of other gateways: A switch-
out probability that determines the number of nodes in
its own tree that should probe other channels, and a
switch-in probability that determines the channel that
those switched-out nodes should probe. The gateway
then disseminates these probabilities to its tree’s nodes.
Upon receiving the command, each node probabilisti-
cally probes other channels and switches to the channel
that provides the best parent link.
We report results from one of the production deploy-
ments. This deployment consists of 694 Genomotes (in-
cluding 174 wireless masters) in a 12,000 sq-ft Microsoft
colocation facility. Slave Genomotes sample their tem-
perature and humidity sensors every 30 seconds. The
network uses up to four wireless channels. The system
has been running for more than 7 months and collects
more than 2.5 million measurement records per day.
Figure 2 presents the data yield over a 72-hour period.
The percentage on the Y-axis is the ratio between the
0 12 24 36
48 60 72
Data Yield (%)
Figure 2: Average hourly data yield over a 72-hour pe-
riod from the 694 sensors in the production deployment.
< 10< 20 < 30
Data Latency (sec)
< 40< 50< 60 More
Figure 3: Data collection latency distribution of 10,000
data samples. The network had 3 gateways.
actual data received at the gateway during that hour
and the expected amount of data, based on the sampling
rate (i.e., 120 records per hour). One can see that the
network has a data yield close to 100%. The minor data
loss was the result of a misconfiguration where the base
stations immaturely gave up on a record after only a few
failed tries. Figure 3 shows that over 90% data have a
latency of less than 30 seconds.
Last but not least, a considerable amount of effort
was necessary to deploy thousands of sensors. One ma-
jor pre-deployment overhead we have observed is assign-
ing unique sensor addresses as this requires bookkeeping
and multiple versions of the software that differ in only
a single variable. To streamline the process, we include
a 64-bit serial ID chip and a matching barcode on each
Genomote. Since the collected measurements have spa-
tial significance, the barcode minimizes human error at
all stages of the deployment.
 C. L. Belady. In the data center, power and cooling costs more
than the it equipment it supports. ElectronicsCooling magazine,
3(1), February 2007.
 O. Gnawali, R. Fonseca, K. Jamieson, and P. Levis. Robust and
efficient collection through control and data plane integration.
Technical Report SING-08-02, 2008.
 C.-J. M. Liang, J. Liu, L. Luo, and A. Terzis. Racnet: Reliable
acquisition network for high-fidelity data center sensing.
Technical Report MSR-TR-2008-144, Microsoft Research, 2008.
 J. Liu, B. Priyantha, F. Zhao, C.-J. M. Liang, Q. Wang, and
S. James. Towards fine-grained data center cooling monitoring
using racnet. In HotEmNets, 2008.
 R. Musaloiu-E., C.-J. M. Liang, and A. Terzis. Koala: Ultra-Low
Power Data Retrieval in Wireless Sensor Networks. In IPSN,
 The Green Grid. The green grid data center power efficiency
metrics: PUE and DCiE, 2007.