With the adoption of the Energy Efficient Ethernet (EEE) standard, latency will increase. This can compromise the use of EEE switches in low latency applications, such as High Performance Computing. Here's, a simple technique to reduce the latency in store and forward switches showing that it can effectively mitigate the effects of EEE in switch latency.
The success of Ethernet as a dominant technology for wired Local Area Networks (LANs) has motivated an increased interest to use Ethernet technology in environments that require low latencies. With the adoption of the new Energy Efficient Ethernet (EEE) standard in coming generations of Ethernet switches, latency will increase. This can compromise the use of EEE switches in low latency applications, such as High Performance Computing. In this article, a simple technique to reduce the latency in store and forward switches is presented showing that it can effectively mitigate the effects of EEE in switch latency.
The wide adoption of Ethernet technology in LANs has enabled economies of scale, thus reducing the cost of Ethernet equipment when compared with other alternative technologies. This in turn has motivated an increased interest to use Ethernet in applications for which it was not originally intended, such as real-time industrial systems [1] and high performance computing clusters [2]. In these new applications, low latency is critical and special care must be taken, both in the network topology and in frame transmission ordering, to ensure that latency requirements are met [1].
The IEEE 802.3az (Energy Efficient Ethernet) task force is currently working on a new Ethernet standard that will greatly reduce the energy consumption of Ethernet transceivers [3]. Additional energy savings can be obtained in other elements of the switches when EEE is implemented [4]. The EEE standard proposes the use of a low power mode when there are no frames to transmit. This allows most of the transceivers elements to be put into a quiet state, greatly reducing the energy consumption. When new frames arrive, the transceivers are awakened and the link is activated such that transmission can take place. Additionally, periodic refresh intervals are scheduled when the transceiver is in low power mode to ensure that the receiver elements are aligned with the current status of the channel.
The overall operation of EEE is illustrated in Figure 1. An initial evaluation of the expected performance of EEE in terms of Energy savings shows that significant savings can be obtained when the links operate at low loads [5].
As shown in Figure 1, when a frame arrives for transmission and the link is in the low power mode, it has to wait until the link is activated before transmission can begin. This introduces additional latency which is on the order of microseconds. Table I shows this additional latency for different link speeds as specified in [6]. This latency increase will not be an issue for most LANs configurations, but can compromise the performance in low latency environments such as real time industrial systems and high performance computer clusters. In this work, a technique to reduce this additional latency increase in Ethernet switches is presented.
Early Destination Lookup
In a switch, the use of EEE will cause additional latency when the frame's destination port is in low power mode. Today most Ethernet switches are store and forward. That is, they entirely receive a frame before forwarding it. For those switches, the destination port will be determined after the frame is completely received and only then the outgoing port will be activated. To reduce latency, we propose an early destination lookup once the destination address is received to obtain the destination port before the frame has been completely received. This is possible because the destination address is at the beginning of the Ethernet Frame as illustrated on Figure 2.
If the destination port is in low power mode, activation can take place while the rest of the frame is being received in the incoming port. This is similar to cut-through switching, in which an early destination lookup is performed so that frame transmission can start on the outgoing link while the rest of the frame is being received [7] . However, in the proposed approach, the early lookup is only used to wake the outgoing link, so that additional frame processing on other parts of the frame can be done before a final forwarding decision is made. If a decision is made to not forward the frame, for example, due to security reasons, the only cost would be a small amount of energy wasted on waking up the outgoing link. However, if the frame is forwarded as normal, then the latency required to wake the destination overlaps with frame reception and is reduced or eliminated.
To assess the benefits of the proposed approach, frame sizes of 60 bytes and 1500 bytes will be considered. They correspond to small and large frames, and represent the two extreme cases in terms of latency and benefits of the proposed solution. Coincidently these frames are commonly found in LAN traffic as shown in Figure 3. It will also be assumed that the early destination lookup will enable us to start link activation once the first six bytes of the frame (containing the destination address) are received.
