Ece.uci.edu

Distributed Embedded Systems for Low Power:
A Case Study
University of California, Irvine, CA 92697-2625, Abstract
scheduling techniques. As DVS reaches its limit on a singleprocessor, researchers turn to multiple processors to create A multiple-processor system can potentially achieve higher energy savings than a single processor, because Multiple processors can potentially achieve higher en- the reduced workload on each processor creates new op- ergy savings than a single processor. By partitioning the portunities for dynamic voltage scaling (DVS). However, workload onto multiple processors, each processor is now as the cost of communication starts to match or surpass responsible for only a fraction of the workload and can op- that of computation, many new challenges arise in making erate at a lower voltage/frquency level with quadratic power DVS effective in a distributed system under communication- saving. Meanwhile, the lost performance can be compen- intensive workload. This paper discusses implementation sated by the increased parallelism. Another advantage with issues for supporting DVS on distributed embedded pro- a distributed scheme is that heterogeneous hardware such as cessors. We implemented and evaluated four distributed DSP and other accelerators can further improve power effi- schemes: (1) DVS during I/O, (2) partitioning, (3) power- ciency of various stages of the computation through special- failure recovery, and (4) node rotation. We validated the re- ization. Although a tightly-coupled, shared-memory multi- sults on a distributed embedded system with the Itsy pocket processor architecture may have more power/performance computers connected by serial links. Our experiments con- advantages, they are not as scalable as distributed, message- firmed that a distributed system can create new DVS op- portunities and achieve further energy savings. However, While distributed systems have many attractive proper- a surprising result was that aggregate energy savings do ties, they pay a higher price for message-passing communi- not translate directly into a longer battery life. In fact, the cations. Each node now must handle not only I/O with the best partitioning scheme, which distributed the workload external world, but also I/O on the internal network. Pro- onto two nodes and enabled the most power-efficient CPU gramming for distributed systems is also inherently more speeds at 30–50%, resulted in only 15% improvement in difficult than for single processors. Although higher-level battery lifetime. Of the four techniques evaluated, node ro- abstractions have been proposed to facilitate distributed tation showed the most measurable improvement to battery programming, these abstraction layers generate even more lifetime at 45% by balancing the discharge rates among the inter-processor communication traffic behind the scenes.
While this may be appropriate for high-performance clus-ter computers with multi-tier, multi-gigabit switches likeMyrinet or Gigabit Ethernet, such high-speed, high-powercommunication media are not realistic for battery-powered Introduction
embedded systems. Instead, the low-power requirementhave constrained the communication interfaces to much Dynamic voltage scaling (DVS) is one of the most stud- slower, often serial interfaces such as I2C and CAN. As a ied topics in low-power embedded systems.
result, even if the actual data workload is not large on an CMOS characteristics, the power consumption is propor- absolute scale, it appears expensive relatively to the com- tional to V 2; while the supply voltage V is linearly propor- putation performance that can be delivered by today’s low- tional to the clock frequency. To fully exploit such quadratic power vs. voltage scaling effects, previous studies have ex- The effect of I/O on embedded systems has not been tensively explored DVS with real-time and non-real-time well studied in existing DVS works. Many existing DVS techniques have shown impressive power savings on a sin- We also use Itsy’s on-board power instrumentation features gle processor. However, few results have been fully qual- to collect data for the power characteristics. Our results ified in the context of an entire system. Even fewer have confirmed that the distributed DVS scheme combined with been validated on actual hardware. One common simplify- efficient load balancing by rotating the nodes achieved the ing assumption is to ignore I/O. Embedded systems (includ- highest measured energy saving and extended the battery ing single-processor systems) that perform no I/O are not realistic. I/O can actually enhance computation by creat-ing opportunities for DVS through parallelism. At the sametime, I/O can also compete with computation for time and Related Work
power budgets, thereby lowering the limit on power savingsachievable by DVS. The effects of I/O on DVS is not yetwell understood, and the problem is further complicated bythe trend towards DVS in distributed systems.
Real-time scheduling has been extended to DVS The contributions of this paper are two-fold.
scheduling on variable-voltage processors.
we demonstrate the gap between CPU-centric DVS claims scheduling model was introduced by Yao et al [10], then ex- and actual attainable power savings by implementing tended and refined by Yasuura [6], Quan [7] and many stud- a full-featured distributed embedded system running a ies in variations for real-time scheduling problems. Since communication-bound, communication-intensive workload power is a quadratic function of the supplying voltage, low- with expensive I/O. This work also contrasts with sensor ering the voltage can result in significant savings while networks, which may be distributed, networked, and low- still enabling the processor to continue making progress power, but they are 99% idle, perform very little computa- such that the tasks can be completed before their deadlines.
tion and communication, and are soft real-time. Our case These techniques often focus on the energy reduction to the study considers much higher computation workload under processor only; while the power consumption of other com- tight timing constraints. Without much slack, DVS cannot ponents, including memory, I/O, is ignored. The results are be very effective, but the expensive I/O turns out to be a new DVS has been applied to benchmark applications such Our second contribution is the set of principles and pit- as JPEG and MPEG in embedded systems. Im et al [4] pro- falls in global power optimization. Our findings confirmed poses to buffer the incoming tasks such that the idle period that parallelism can indeed create new opportunities for between task arrivals can be utilized by DVS. Shin et al [8] DVS to achieve further energy savings; however, one must introduces an intra-task DVS scheme that maximally uti- avoid many pitfalls in order to achieve these savings on a lizes the slack time within one task. Choi et al [2] presents distributed architecture powered by separate batteries. A a DVS scheme in an MPEG decoder by taking advantage of single battery failure can be disastrous to the entire sys- the different types of frames in the MPEG stream. These tem. We observed that global energy optimization can often techniques can be validated with measurement on real or contradict the goal to maximize the uptime of a distributed, emulated platforms. However, they are also computation- battery-powered system. This is due to the fact that global oriented such that the processor performs only very little, if optimization does not guarantee a locally near-optimal con- any I/O. The impact of I/O still remains under-studied.
figuration for each distributed node. An ill-configured node DVS has recently been extended to multi-processor sys- operating at an energy-inefficient point can drain its battery tems. Weglarz [9] proposes partitioning the computation quickly and bring down the whole system. Our experiments onto a multi-processor architecture that consumes signif- indicated load balancing is one of the key factors in decid- icantly less power than a single processor.
ing the uptime of the first faulting node. Special considera- fundamental difference in applying techniques for multi- tions, including partitioning, scheduling, synchronization, processors to distributed systems. Minimizing the global load balancing and power failure detection and recovery, energy consumption will extend the battery life only if the must be carefully coordinated with DVS, or else the same whole system is assumed to be powered by a single battery DVS techniques will be counterproductive.
unit. In a distributed environment, each node is power by This paper first reviews DVS techniques, the application a dedicated battery. Even a globally optimal solution may example, and the experimental platform. We chose the Itsy cause poor battery efficiency locally and result in shortened pocket computer as our experimental platform: it supports system uptime as well as loss of battery capacities. Maleki a rich, well-documented set of DVS routines, and it is also et al [5] analyzes the energy efficiency of routing proto- available to other researchers who wish to reproduce these cols in an ad-hoc network and shows that the global optimal results. Because Itsy runs Linux, it is easy to port full- schemes often contradicts the goal to extend the lifetime of fledged, distributed programs and experimental tools to it.
Figure 1. Block diagram of the ATR algorithm.
Figure 2. The timing vs. power diagram of a
single node.
Motivating Example
Figure 3. The timing vs. power diagram of two
nodes.
We select an image processing algorithm, automatic tar- get recognition (ATR) as our motivating example to evalu-ate a realistic application under I/O pressure. Its block di-agram is shown in Fig. 1. The algorithm is able to detectpre-defined targets on an input image. For each target, aregion of interest is extracted and filtered by templates. Fi-nally, the distance of each target is computed. A throughput When multiple nodes are configured as a distributed sys- constraint is imposed such that the image frames must be tem, we organize them in a pipeline for the ATR algorithm.
Fig. 3 shows the timing vs. power diagram of two pipelined We evaluate a few DVS schemes on one or more em- nodes performing the ATR algorithm. Node1 maps first two bedded processors that perform the ATR algorithm. Un- function blocks, and Node2 performs the other two blocks.
like many DVS studies that ignore I/O, we assume that all Node1 receives one frame from the data source, and it pro- nodes in our system are connected to a communication net- cesses the data and sends the intermediate result to Node2 work. This network carries data from external sources (e.g., in D seconds. After Node2 starts receiving from Node1, it a camera, sensor, etc.) internal communications between finishes its share of computation and sends the final result the nodes, and to an external destination (e.g., a PC). This to the destination within D seconds. Fig. 3 shows that if the study assumes only one image and one target are processed data source keeps producing one frame every D seconds, at a time, although a multi-frame, multi-target version of the and both Node1 and Node2 can send their results also in D seconds, then the distributed pipeline is able to provide one We refer to each embedded processor as a node. A node result in every D seconds to the destination.
is a full-fledged computer system with a voltage-scalableprocessor, I/O devices, and memory. Each node performs We use generic TCP/IP sockets to implement reliable a computation task PROC and two communication tasks communication, although it could be further optimized. We RECV and SEND. RECV receives data from the external believe this is reasonable and is much lighter weight than source or another node. The data is processed by PROC, middleware such as CORBA, which some researchers advo- which consists of one or more functional blocks of the ATR cate even on small devices. We also assume the workload of algorithm. The result is transmitted by task SEND to an- the algorithm is fixed such that DVS opportunities are lim- other node or the destination. Due to data dependencies, ited by I/O and timing constraints. Our purpose in this study tasks RECV , PROC, and SEND must be fully serialized for is to explore the computation-I/O interaction and their im- each node. In addition, they must complete within a time pact on DVS, but not to specifically optimize for either I/O period called the frame delay D, which is defined as the per- power or computation power. Other techniques that reduce formance constraint. Fig. 2 illustrates the timing vs. power communication or computation power under variable work- load can be readily brought into the context of this study.
Figure 5. Networking multiple Itsy units with
a host computer.
Figure 4. The block diagram of Itsy [1].
Experimental Platform
transparently as if they were on the same TCP/IP network.
The network configuration is shown in Fig. 5.
We use the Itsy pocket computers as distributed nodes, The serial link might not be the best choice for inter- connected by a TCP/IP network over serial links.
connect, but it is often used in real life due to power con- present the performance and power profiles of the ATR al- straints. A high-speed network interface requires several gorithm running on Itsy computers and define the metrics Watts of power, which is too high for battery-powered em- In this paper, our primary goal is to investigate the The Itsy Pocket Computer
new opportunities for DVS-enabled power vs. performancetrade-offs in distributed embedded systems with intensive The Itsy pocket computer is a full-fledged miniaturized I/O. Given the limitations of serial ports, we do not intend computer system developed by Compaq Western Digital to propose our experimental platform as a prototype of a Lab [1, 3]. It supports DVS on the StrongARM SA-1100 new distributed network architecture. We chose this net- processor with 11 frequency levels from 59 – 206.4 MHz work platform primarily because it represents the state of over 43 different voltage levels. Itsy also has 32MB flash the art in power management capabilities. It is also consis- memory for off-line storage and 32MB DRAM as the main tent with the relatively expensive communication, both in terms of time and energy, seen by such systems. We ex- lithium-ion battery pack. Due to the power density con- pect that our findings in this paper can be applied to many straint of the battery, Itsy currently does not support high- communication-intensive applications on other network ar- speed I/O such as Ethernet or USB. The applicable I/O ports chitectures, where communication is a key factor for both are a serial port and an infra-red port. Itsy runs Linux with networking support. Its block diagram is shown in Fig. 4.
Performance Profile of the ATR Algorithm
Network Configuration
Each single iteration of the entire ATR algorithm takes We currently use the serial port as the network interface.
1.1 seconds to complete on one Itsy node running at the We set up a separate host computer as both the external peak clock rate of 206.4 MHz. When the clock rate is re- source and destination. It connects the Itsy nodes through duced, the performance degrades linearly with the clock multiple serial ports established by USB/serial adaptors. We rate. The PPP connection on the serial port has a maximum setup individual PPP (point-to-point protocol) connections data rate of 115.2 Kbps, though our measured data rate is between each Itsy node and the host computer. Therefore roughly 80 Kbps. In addition, the startup time for establish- the host computer acts as the hub for multiple PPP net- ing a single communication transaction takes 50–100 ms.
works, and it assigns a unique IP address to each Itsy node.
The computation and communication behaviors are profiled Finally, we start the IP forwarding service on the host com- and summarized in Fig. 6. The functional blocks can be all puter to allow Itsy nodes to communicate with each other combined into one node or distributed onto multiple nodes long delays thus consume a significant amount of energy, although the communication power level is not the highest.
As a result, I/O energy becomes a primary target to optimizein addition to DVS on computation.
We evaluate several DVS techniques by a series of ex- periments with one or more Itsy nodes. A baseline config-uration is a single Itsy node running the entire ATR algo- Figure 6. Performance profile of ATR on Itsy.
rithm at the highest clock rate. It is able to produce oneresult in every D seconds. For all experiments, we fix this frame delay D as the performance constraint and keep theItsy node(s) running until the battery is fully discharged.
The energy metric can be measured by the battery life T (N)when N nodes with N batteries are being used. The com- pleted workload F(N) is the number of frames completedbefore the battery exhaustion. The battery life in the base- Current (mA)
line configuration is T (1). Since the frame delay D is fixed, the host computer will transmit one frame to the first Itsynode in every D seconds. The Itsy node(s) are also able to complete and send one result back to the host in every Dseconds. In an N-node pipeline, there is a pipeline startup delay (N − 1) × D before the first result can be produced.
Therefore, T (N) = F(N) × D + (N − 1) × D. Since F(N) Freq (MHz)
is at least a few thousand frames in our experiments while N = 2, the pipeline startup overhead is ignored such that Communication
Computation
T (N) = F(N) × D.
The battery life T (N) is also called the absolute battery Figure 7. Power profile of ATR on Itsy.
life. We also define the normalized battery life Tnorm(N) =T (N)/N to quantify the energy savings for fair compar-isons. The rationale behind this distinction is that, the to- in a pipeline. In the single node case there are no communi- tal lifetime of N batteries should be at least N times that of cations between adjacent nodes, although the node still has a single battery, or else they are less energy efficient. For to communicate with the host computer.
example, a two-node system with two batteries should lastat least twice as long as a single node does. To make com- Power Profile of the ATR Algorithm
parisons easier, we define the normalized battery life ratioRnorm(N) = Tnorm(N)/T (1). In the baseline configuration, Fig. 6 shows the net current draw of one Itsy node. The Tnorm(1) = T (1), Rnorm(1) = 100%.
horizontal axis represents the frequency and correspond-ing voltage levels. The data are collected by Itsy’s built-in Techniques under Evaluation
power monitor. During all experiments the LCD screen andthe speaker are turned off to reduce unnecessary power con- We first define the baseline configuration as a reference sumption. The execution of the ATR algorithm on Itsy has to compare experimental results. We briefly review the DVS three modes of operations: idle, communication and com- techniques to be evaluated by our experiments.
putation. In idle mode, the Itsy node has neither I/O nor anycomputation workload. In communication mode, it is eithersending or receiving data through the serial port. In com- Baseline Configuration
putation mode, it executes the ATR algorithm. Fig. 6 showsthe three curves range from 30 mA to 130 mA, indicating a The baseline configuration is a single Itsy node perform- power range from 0.1W to 0.5W. The computation always ing the entire ATR algorithm. It operates at the highest CPU dominates the power consumption. However, due to the clock rate of 206.4 MHz. The processing task PROC re- slow data rate of the serial port, communication tasks have quires 1.1 seconds to complete. The node also needs 1.1 and nodes are allowed to run at much lower clock rates. The second and third schemes have excessive internal commu- nication. Therefore computation must run faster, otherwise they cannot produce the results in D = 2.3 seconds. Espe- cially in the third scheme, Node1 is not capable of complet- ing its work on time unless clocked at 380 MHz, which ex- ceeds the maximum clock rate. We choose the first scheme for all distributed DVS experiments, although the compu- Figure 8. Three partitioning schemes.
tation workload is still unbalanced. However, it is the op-timal partitioning between computation and I/O in a sensethat Node1 also takes more than 90% of the total commu- 0.1 seconds to receive and send data, respectively. There- nication payload in addition to its 10% share of the total fore the total time to process one frame is D = 2.3 seconds.
Based on the metrics we defined in Section 4.5, we fix thisframe delay D = 2.3 seconds in all experiments.
Distributed DVS with Power Failure Recovery
DVS during I/O
In general, it is impossible to evenly distribute the work- load to each node in a distributed system. In many cases The first technique is to perform DVS during the I/O pe- even the optimal partitioning scheme yields very unbal- riod. Since the application is tightly constrained on timing anced workload distribution. In our experiments, Node2 with expensive I/O delay, there is not much opportunity for with more workload will have to run faster thus its battery DVS on computation without a performance penalty. On the will exhaust sooner. After one node fails, the distributed other hand, since the Itsy node spends a long time on com- pipeline will simply stall although the remaining nodes still munication, it is possible to apply DVS during I/O. Based have sufficient battery capacity to keep working. This will on the power characteristics shown in Fig. 7, I/O can operate result in unnecessary loss of battery capacity.
at a significantly low-power level at the slowest frequency One potential solution is to recover from the power fail- ure on one node by detecting the faulting node dynami-cally and migrating its computation to neighboring nodes.
Distributed DVS by Partitioning
Such techniques normally require additional control mes-sages between nodes, thereby increasing I/O pressure on the Partitioning the algorithm onto multiple nodes can cre- already I/O-bound applications. Since these messages will ate more time per slot for DVS on each distributed node.
also cost time, they will force an increase of computation However, since the application is already I/O-bound, addi- speed such that the node will fail even sooner.
tional communication between nodes can further increase As a proof of concept, we implement a fault recovery the I/O pressure. A few concerns must be taking into ac- scheme as follows. Each sending transaction must be ac- count to correctly balance computation and I/O. First, each knowledged by the receiver. A timeout mechanism is used node must be able to complete its tasks RECV , PROC, and on each node to detect the failure of the neighboring nodes.
SEND within D = 2.3 seconds. With an unbalanced parti- The computation share of the failed node will then migrate tioning, a node can be overloaded with either excessive I/O to one of its neighboring nodes. The message reporting or heavy computation, such that it cannot finish its work on a faulting node can be encapsulated into the sending data time and then the whole pipeline will fail to meet the perfor- stream and the acknowledgment. Therefore, the informa- mance constraint. Second, additional communication can tion can be propagated to all nodes in the system. As men- potentially saturate the network such that none of the nodes tioned in Section 4.3, the acknowledgment signal requires can guarantee to finish their workload on time. Finally, the a separate transaction, which typically costs 50–100 ms in distributed system should deliver an extended battery life in addition to the extended I/O delay. Since the frame delay the normalized term, not just a longer absolute uptime.
D is fixed, the processor must run faster to meet the tim- We experiment with two Itsy nodes, although the results ing constraint due to the increased I/O delay to support the do generalize to more nodes. Based on the block diagram in Fig. 6, three partitioning schemes are available and il-lustrated in Fig. 8. The first scheme, where Node1 is only Distributed DVS with Node Rotation
responsible for target detection and Node2 performs the re-maining three functional blocks, is clearly the best among As an alternative to the power failure recovery scheme all three solutions. Due to the least amount of I/O, both in Section 5.4, we balance the load on each node more ef- ficiently with a new technique. If all nodes are evenly bal- ready saved communication energy by eliminating a pair of anced, after the first battery fails, then the other batteries communication transactions. Therefore the energy cost of will also exhaust shortly. There is not much battery capacity the transition is also minimal, if not zero. For brevity the to enable the remaining nodes to continue making progress, illustration is omitted for N > 2.
even if a recovery scheme allows them to. Therefore, powerfailure recovery is not necessary and its expensive overhead Experimental Results
can be avoided, while the battery capacity can be still uti-lized efficiently.
We evaluate the DVS techniques described in Section 5 We designed a new load balancing technique called node by experiments then analyze the results in the context of a rotation. The idea is that if we can shuffle the workload on all nodes, such that the lightly-loaded nodes will have moreworkload and the heavily-loaded nodes can “rest,” then the (0A, 0B) Initial Evaluation without I/O
workload on each node will be evened out after a few shuf-fles. However, reconfiguring the nodes in a pipeline gener- Before experimenting DVS with I/O, we first perform ally requires a pipeline stall (or flush) followed by a restart, two simple experiments on a single Itsy node to explore the which will incur both performance and energy penalties.
potential of DVS without I/O. The single Itsy node reads Our node rotation scheme involves minimal overhead, and local copies of the raw images and it only computes the re- it works as follows. At a given moment, each node in the sults, instead of receiving images from the host and send- pipeline will perform the following procedure.
ing the results back. Therefore there is no communication After Nodei, for i = 1, 2,. , N − 1, finishes the pro- delay or energy consumption involved. (0A): We use one cessing task PROCi, it will not send the result to the Itsy node to keep running the entire ATR algorithm at the next node Nodei+1. Instead Nodei reconfigures itself to full speed 206.4 MHz. Its battery will exhaust in 3.4 hours Nodei+1. That is, it continues performing the processing with 11.5K frames completed. (0B): We setup the second task PROCi+1 of Nodei+1, with the input data already avail- Itsy node to execute at the half speed 103.2 MHz. Then able from the result of PROCi. After each Nodei finishes it is able to continue operating for 12.9 hours by finishing PROCi+1, then it sends the result to Nodei+1 (that has been 22.5K frames. At the half clock rate, the Itsy computer can reconfigured as Nodei+2) except for node NodeN−1 (that complete twice workload as much as it can do at the full has been reconfigured as the last node NodeN) that will send the final result to the host. Afterwards, each node Node i will We overload the metrics notation we defined in Sec- tion 4.5 as follows: T maps the experiment label to the The last node NodeN will reconfigure itself as the first total battery life, and F maps the experiment label to the node Node1, and will start receiving from the host, process- number of frames processed. Here, T (0A) = 3.4 (hours), ing the data with PROC1 and sending the result to the next F(0A) = 11500. T (0A) = 12.9, F(0B) = 22500. Note that node. During such a procedure, the last node is rotated to these results are not to be compared with other experiments, the front of the pipeline. If rotation is performed once in since there is no communication and no performance con- every certain number of frames, after N rotations, the work- load on each node is evenly balanced.
The results are promising for having more nodes as a Fig. 9 illustrates this procedure. After Node1 finishes distributed system. By using two Itsy nodes running at the PROC1 for the Ith frame, it will continue on PROC2 then half speed, the system should be able to deliver the same send the Ith result to the host. Then it “becomes” Node2.
performance as one Itsy node at the full speed does, while Meanwhile Node2 becomes Node1 such that it will receive completing four times the workload by using two batteries.
the (I + 1)th frame from the host, process it by PROC1 and However, such an upperbound can only be achieved without pass the intermediate result to Node1 (that has already be- come Node2). During the transition period, Node1 elim-inates one SEND transaction and so does Node2 a RECV (1) Baseline configuration
transaction. This extra idle time slot is previously allocatedfor a long-delay communication transaction. It should be We defined the baseline configuration in Section 5.1.
sufficient for both nodes to load the new code into memory The single Itsy node running at 206.4 MHz can last for 6.13 and reconfigure themselves as each other. There is no per- hours and finish 9.6K frames before the battery dies. That formance loss since the host can still send one frame and is T (1) = Tnorm(1) = 6.13, F(1) = 9600, Rnorm(1) = 100%.
receive one result in every D seconds, thus the throughput Compared with experiment (0A) without I/O, the completed of the pipeline remains the same. Both nodes must consume workload is 17% less since the node must spend a long time some energy to refresh their code memory but they have al- Idle time for
reconfiguration
Idle time
for
reconfi-
guration
Figure 9. Node rotation on two nodes.
(1A) DVS during I/O
(2) Distributed DVS by Partitioning
As Section 5.2 suggests, we apply DVS to I/O periods, Since there is no further opportunities for DVS with the such that during sending and receiving the Itsy node oper- single node, from now we evaluate distributed configura- ates at 59 MHz, while in computation it still runs at 206.4 tions with two Itsy nodes in a pipeline. In Section 5.3 we se- MHz. From our measurement communication delay does lected the best partitioning scheme, in which two Itsy nodes not increase at a lower clock rate. Thus the performance re- operate at 59 MHz and 103.2 MHz, respectively. The dis- mains the same as D = 2.3 seconds. Through DVS during tributed two-node pipeline is able to complete 22.1K frames I/O, the battery life is extended to 7.6 hours and it is able in 14.1 hours. That is, T (2) = 14.1, F(2) = 22100. Com- to finish 11.9K frames. That is T (1A) = Tnorm(1A) = 7.6, pared to experiment (1), the battery life is more than dou- F(1A) = 11900, Rnorm(1A) = 124%, indicating a 24% in- bled. However, after normalizing the results for two batter- ies, Tnorm(2) = 7.05, Rnorm(2) = 115%, meaning the batterylife is only effectively extended by 15%. Distributed DVS Note that F(1A) > F(0A) = 11500. Even though the is even less efficient than (1A), in which DVS during I/O Itsy node is communicating a large amount of data with the can extend 24% of the battery capacity.
host computer, it completes more workload than it does inexperiment (0A) without I/O. This is due to the recovery ef- There are a few reasons behind the results. First, when fect of batteries. Recovery effect indicates that if a battery Node2 fails, the pipeline simply stalls while plenty of en- continues experiencing a high discharge current, it capacity ergy still remains on the battery of Node1. Second, Node2 will exhaust sooner, as the results of (0A) and (1) show. On always fails first because the workload on the two nodes is the other hand, if the discharge current can drop to a lower not balanced very well. Node2 has much more computation level, the lost capacity can be partially recovered. In this load and it has to run at 103.2 MHz; while Node1 has very experiment (1A), the current level is reduced from 110 mA little computation such that it operates at 59 MHz. How- to 40 mA (Fig. 7) for 1.2 second in every 2.3 seconds. This ever, this partitioning scheme has already been optimal with allows the battery to “rest” after heavy discharge on com- the maximally balanced load. If we choose other partition- putation and recover its capacity. As a result, the battery ing schemes, the system will fail even sooner as analyzed in (2A) Distributed DVS during I/O
73.7 MHz, and Node2 at 118 MHz. We also perform DVSduring I/O for both nodes. The result is, T (2B) = 15.72, DVS during I/O (1A) can extend 24% battery life for a F(2B) = 24500, Tnorm(2B) = 7.86 and Rnorm(2B) = 128%.
single node. We expect the distributed pipeline can also With our recovery scheme, the system can last longer benefit from the same technique by applying DVS during than (2) and (2A). However there is no significant im- I/O for distributed nodes. Among the two Itsy nodes, Node1 provement compared to the simple DVS during I/O scheme is already configured to the lowest clock rate. Therefore, (1A). Since both nodes must run faster, Node2 will fail we can only reduce the clock rate of Node2 to 59 MHz more quickly after completing 19.5K frames and Node1 can during its I/O period and leave it at 103.2 MHz for com- pick up another 5K frames until all batteries have exhausted.
putation. The result is T (2A) = 14.44, F(2A) = 22600, Power failure recovery allows the system to continue func- tioning with failed nodes. However it is expensive in a sense norm(2A) = 7.22 and Rnorm(2A) = 118%. Only 3% more battery capacity is observed comparing with experiment (2).
that it must be supported with additional, expensive energy Distributed DVS during I/O is not as effective as DVS during I/O for a single node. According to the power profilein Fig. 7, from (1) to (1A) the discharge current drops from (2C) Distributed DVS with Node Rotation
110 mA to 40 mA during I/O periods, which take the halfof the execution time of the single node. However, from (2) Up to now the distributed DVS approaches do not seem to (2A), we only optimize for Node2 that has already op- effective enough. In experiment (2) and (2A), the failure of erated at a low-power level during I/O (55 mA). By DVS Node2 shuts down the whole system. Experiment (2B) al- during its I/O periods, the discharge current decreases to lows the remaining Node1 to continue. However the power 40 mA. Thus, the 15 mA reduction is not as considerable failure recovery scheme also consumes energy before it can compared with the 70 mA saving in experiment (1A). In ad- save energy. What prevents a longer battery life is the un- dition, Node2 does not spend a long time during I/O. It only balanced load between Node1 and Node2. In this new ex- communicates 700 Bytes in very short periods. Therefore, periment we implemented our node rotation technique pre- the small reduction to a small portion of power use con- sented in Section 5.5, combined with DVS during I/O. Since tributes trivially to the system. On the other hand, Node1 there is no performance penalty, two nodes can still operate has heavy I/O load. However, since it runs at the lowest at at 59 MHz and 103.2 MHz. By node rotation in every 100 power level, there is no chance to further optimize its I/O frames, the battery life can be extended to T (2C) = 17.82, F(2C) = 27900, Tnorm(2C) = 8.91 and Rnorm(2C) = 145%.
From experiments (2) and (2A) we learn a few lessons.
This is the best result among all techniques we have eval- Although there are more distributed DVS opportunities uated. Node rotation allows the workload to be evenly dis- whereas not available on a single processor, the energy sav- tributed over the network thus maximally utilizes the dis- ing is no longer decided merely by the processor speed. In a tributed battery capacity. There is also an additional benefit.
single processor, minimizing energy directly optimizes the Since both nodes alternate their frequency between 103.2 life time of its single battery. However in a distributed sys- MHz and 59 MHz, both batteries can take advantage of the tem, batteries are also distributed. Minimizing global en- recovery effect to further extend their capacity.
ergy does not guarantee to extend the lifetime for all bat- To summarize, our experimental results are presented in teries. In our experiments, the load pattern of both commu- Fig. 10. Both absolute and normalized battery lives are il- nication and computation decides the shortest battery life, lustrated, with normalized ratios annotated. The results of which often determines the uptime of the whole system.
experiments (0A) and (0B) without communication are notincluded since it is not proper to compare them with I/O- (2B) Distributed DVS with Power Failure Re-
bound results. It should be noted that the effectiveness of these techniques is application-dependent. Although exper-iment (2) and (2A) do not show much improvement in thiscase study, the corresponding techniques can still be effec- In experiments (2) and (2A), the whole distributed sys- tem fails after Node2 fails, although Node1 is still capableof carrying on the entire algorithm. We attempt to enablethe system to detect the failure of Node2 and reconfigure Conclusion
the remaining Node1 to continue operating. Our approachis described in Section 5.4. We use the same partition- This paper evaluates DVS techniques for distributed low- ing scheme in (2) and (2A). Due to the additional com- power embedded systems. DVS has been suggested an ef- munication transactions for control messages, both nodes fective technique for energy reduction in a single processor.
As a result, Node1 must operate at As DVS opportunities diminish in communication-bound, Figure 10. Experiment results.
time-constrained applications, a distributed system can ex- [2] K. Choi, K. Dantu, W.-C. Cheng, and M. Pedram. Frame- pose richer parallelism that allows further optimization for based dynamic voltage and frequency scaling for a MPEG both performance and DVS opportunities. However, the de- decoder. In Proc. International Conference on Computer- signers must be aware of many tricky and often counter- Aided Design, pages 732–737, November 2002.
[3] W. R. Hamburgen, D. A. Wallach, M. A. Viredaz, L. S.
intuitive issues, such as additional I/O, partitioning, power Brakmo, C. A. Waldspurger, J. F. Bartlett, T. Mann, and K. I.
failure recovery and load balancing, as indicated by our Farkas. Itsy: stretching the bounds of mobile computing.
study. We presented a case study of a distributed embed- IEEE COMPUTER, 34(4):28–36, April 2001.
ded application under various DVS techniques. We per- [4] C. Im, H. Kim, and S. Ha. Dynamic voltage scaling tech- formed a series of experiments and measurements on actual nique for low-power multimedia applications using buffers.
hardware with DVS under I/O-intensive workload, which is In Proc. International Symposium on Low Power Electron- typically ignored by many DVS studies. We also proposed a new load balancing technique that enables more aggres- [5] M. Maleki, K. Dantu, and M. Pedram. Power-aware source sive distributed DVS that maximizes the uptime of battery- routing protocol for mobile ad hoc networks. In Proc. Inter-national Symposium on Low Power Electronics and Design, powered, distributed embedded systems.
[6] T. Okuma, T. Ishihara, and H. Yasuura.
Acknowledgment
scheduling for a variable voltage processor. In Proc. In-ternational Symposium on System Synthesis, pages 24–29, This research is sponsored in part by National Sci- ence Foundation under grant CCR-0205712 and DARPA scheduling for real-time systems on variable voltage proces- PAC/C program under subcontract 4500942474 with Rock- sors. In Proc. Design Automation Conference, pages 828– well/Collins. Special thanks to HP Western Research Lab for providing Itsy Pocket Computers and technical assis- [8] D. Shin, J. Kim, and S. Lee. Low-energy intra-task volt- age scheduling using static timing analysis. In Proc. DesignAutomation Conference, pages 438–443, June 2001.
[9] E. F. Weglarz, K. K. Saluja, and M. H. Lipasti. Minimizing References
energy consumption for high-performance processing. InProc. Asian and South Pacific Design Automation Confer- [1] J. F. Bartlett, L. S. Brakmo, K. I. Farkas, W. R. Hamburgen, T. Mann, M. A. Viredaz, C. A. Waldspurger, and D. A. Wal- [10] F. Yao, A. Demers, and S. Shenker. A scheduling model lach. The itsy pocket computer. Technical Report 2000/6, for reduced CPU energy. In IEEE Annual Foundations of COMPAQ Western Research Laboratory, 2000.
Computer Science, pages 374–382, 1995.

Source: http://www.ece.uci.edu/impacct/d_research/d_pub/ipdps04.pdf