【正文】 out $790 million per year in electricity use. While current Ethernet links and switches are mostly 100 Mb/s and 1 Gb/s, we envision that they are likely to evolve to 10 Gb/s in the near future for several reasons including, 1) everdecreasing prices [20], 2) fast adoption by vendors [18], and 3) increasing bandwidth requirements of multimedia applications within households (for example audio/video transfer between storage device and player, and LANbased multiplayer video games). The contributions of this paper are the proposal and explanation of synchronized coalescing and evaluation of its performance tradeoffs and effects on typical Internet traffic and TCP through simulation. The remainder of this paper is organized as follows. Section II reviews EEE and previous work that has been done in policies to control EEE. Section III presents a microlevel study of the power use of SOHO Ethernet switches and the opportunity of powering down individual ponents. Section IV explains the new synchronized coalescing method. Section V is a simulation evaluation of the method. Section VI describes related work. Section VII summarizes, describes future work, and estimates the potential energy savings that could be gained by largescale deployment of the presented methods.II. OVERVIEW OF ENERGY EFFICIENT ETHERNET (EEE) EEE brings the energy consumption of Ethernet links closer to the ideal consumption, which is directly propor tional to the utilization of the link. Estimates show that using EEE in all current 1 Gb/s edge links in both residential and mercial buildings and network equipment links within residences could save about $180 million/year in the . alone [6]. Two modes are defined in EEE。 Active mode and Low Power Idle (LPI) mode. In Active mode the link is poweredon to transmit packets. When there are no more packets to transmit, the link can enter LPI mode in which the physical layer is powered off and elements in the receiver are stopped. The arrival of a packet to the link can result in the link to wake up in a few microseconds to resume packet transmission (Figure 1). In this figure, Ts is the time needed to enter LPI mode and Tw is the time needed to return to Active mode. During the Tw, Ts and Tr periods the link consumes full power, while during Tq only almost 10% of the full power consumption of the link is needed [25]. The refresh cycle of duration Tr is a periodic link activity to maintain the alignment of receiver elements to channel conditions. It can be assumed that the link uses the same power as in Active mode during transitions [25]. The minimum Tw and Ts for 10GBASET links are and μs respectively [12]. The transition times are relatively high pared to the transmission time of μs for a 1500 byte packet at 10 Gb/s.A. Improving the efficiency of EEE The transition times of EEE are large pared to the transmission time of a packet. For instance, if the link wakes up to transmit a single 1500 byte packet, it would spend μs transmitting the packet and μs for transitioning from LPI mode to Active and back. This means that only about 14% of time is dedicated to transmitting the packet and the rest to the transitions. This inefficiency of EEE was first explored by Reviriego et al. [25] in 2009. EEE can be most efficient when packets arrive back to back in bursts. As a result, only one sleep and one wakeup transition is required per burst which makes the percentage of time the interface spends in active mode close to the link utilization. This best case often occurs in the form of file downloads using TCP where large blocks of data are burst onto a link from a server to a client at a high rate. Conversely, the worst case happens when packets arrive with a fixed interarrival time and a spacing greater than the wake and sleep transition times. As a result, one wake and one sleep transition would be required for transmission of each packet resulting in inefficient operation. A closeto worst case traffic scenario occurs when TCP ACKs are being returned from a client to a server. TCP ACKs are typically small packets and are spacedout evenly (Figure 2). The inefficiency of EEE can be reduced by coalescing the outgoing packets into bursts thus decreasing the number of necessary transitions to one per burst. Packet coalescing for EEE is demonstrated and studied in [6]. EEE with packet coalescing is depicted in Figure 3. As shown in this figure, when all the packets in the transmit queue (or buffer) are transmitted and the buffer bees empty, the link is put to LPI mode after a sleep transition which takes Ts time. The packets which arrive thereafter are not transmitted immediately but are buffered into a coalescing buffer. When a maximum time passes from the arrival of the first packet to the coalescing buffer, or the number of buffered packets reaches a predefined maximum, the link exits the LPI mode, which takes Tw time. All the coalesced packets are transmitted in a single burst. Refresh periods are omitted in Figure 3, since Tq Tr.III. SWITCH ENERGY USE AND TRANSITION TIMES To determine the possible energy savings from synchro nization of LPI periods between all ports in a switch, it is necessary to answer the following three questions: 1) Which ponents of the switch can be powered down?2) How much reduction in total switch power use can be achieved by powering down these ponents? 3) What are the required times to transition these po nents from fullypowered to powereddown mode and back? The main ponent of a typical SOHO Ethernet switch is a single CMOS switch chip. DLink DGS1008G, Linksys EG008W, Netgear GS608, and Trendnet TEG S8 are mon examples of such switches (all are Gigabit Ethernet switches). The ?rst uses Vitesse VSC7388 SparXG5 [27], and the rest use Broad BCM5398 chips [5], both of which are switchonachip ICs that include the switching fabric, Ethernet port blocks, interfaces to external CPUs, memories, and the