Energy-Efficient Ethernet (802.3az): Benefits and Gotchas

Energy-Efficient Ethernet is one of the few green networking features that costs nothing to deploy and actually delivers: it puts a link's PHY into a low power idle state between packets, and the link resumes at full speed with no retraining. The savings are modest per port and significant across a campus, and the failure modes are subtle enough that a lot of engineers disable it after one bad experience. Understanding where EEE helps and where it hurts is the difference between a free saving and an intermittent fault nobody can reproduce.

How Low Power Idle Works

802.3az defines Low Power Idle (LPI). When a link has nothing to transmit, the transmitting PHY signals the peer to go quiet, both ends enter a low-power state, and a periodic refresh keeps the link synchronised. When a frame arrives, the PHY wakes, sends it, and goes quiet again. There is no renegotiation and no link flap. The refresh cycle is fast - on the order of microseconds of wake time - which is why EEE is advertised as transparent to traffic.

The distinction matters between three related features that are frequently conflated:

• EEE (802.3az) - PHY-level low power idle on the link, per-port, negotiated via LLDP or auto-negotiation.
• Ultra Low Power Mode (ULP) - a deeper sleep used by some NICs, where the controller may be slow to respond to link events. Not all of 802.3az, and much more disruptive.
• 802.3bt power saving - the PSE side of PoE, which negotiates a lower power allocation when the powered device needs less. A separate mechanism from EEE, but often deployed together.

Where the Savings Actually Are

Per-port savings are in the range of a few hundred milliwatts to about a watt on gigabit copper, and higher on multi-gigabit. That sounds negligible until you multiply it:

48-port access switch, ~0.35 W saved per idle port
= 16.8 W per switch
x 120 switches in a campus
= ~2.0 kW continuous
= ~17.6 MWh per year at 100% duty
= roughly 8-10 tonnes CO2 depending on grid

The savings are largest exactly where links are idle most of the time: access ports serving printers, idle workstations, IP phones outside working hours, and uplinks that carry bursty rather than sustained traffic. They are smallest on a saturated 100G backbone, because the link is rarely idle long enough to matter.

Enabling It

# Cisco IOS-XE
interface GigabitEthernet1/0/10
 energywise      ! platform dependent; lower-level control is usually automatic
! verify
show power inline
show interfaces GigabitEthernet1/0/10 | include Energy

# Arista EOS
switch(config-if-Et1)# ?
! EEE negotiation is typically enabled by default on copper PHYs
show interfaces ethernet 1 | include (ee|Energy)

# Linux host NIC
sudo ethtool --show-eee eth0
sudo ethtool --set-eee eth0 eee on advertise 1000baseT/Full
ethtool --show-eee eth0

# Windows host NIC - properties dialog or PowerShell
Get-NetAdapterAdvancedProperty -Name Ethernet | Where-Object DisplayName -match "Energy|EEE"

On most campus switches, EEE on copper ports is negotiated and on by default. The interesting cases are the ones where it has been turned off deliberately - often for good reason, sometimes because a previous engineer could not diagnose a symptom.

When EEE Causes Problems

The complaints cluster into four groups:

1. Interoperability with older PHYs. A device that mishandles LPI signalling may drop the first frame after a wake, producing sporadic single-packet loss. This shows up as an application timeout that never reproduces under a continuous ping, because a continuous stream keeps the link awake.
2. Latency jitter. Wake time adds a small, variable delay. For ordinary data traffic it is invisible. For workloads with tight tail-latency budgets - financial trading, industrial control, timing-sensitive applications - it is measurable and unwanted. Runt frames and single-packet traffic are the worst case.
3. Link-state and monitoring confusion. Some monitoring tools infer link activity from PHY signalling. In deep low power states the PHY may hold the link up while the media is quiet, which is correct but surprises tooling that treats "no signal transitions" as a fault.
4. Multi-gigabit and cable quality. Higher-rate copper links have less margin. Marginal cabling plus LPI wake transitions can tip a link from reliable to intermittent. This is one of the few cases where the right fix is the cable, not the feature - the same link-certification discipline described in this structured cabling certification guide applies.

Deciding Per Port Class

A workable policy rather than a blanket setting:

Whichever way you go, make it an explicit configuration standard rather than an inherited default. The expensive version of this feature is the one nobody knows is enabled on a link that intermittently misbehaves.

Measuring the Result

Do not trust the marketing number. Measure:

# switch power draw at the PDU, before and after a change window
# if you have SNMP on the PDU or UPS, trend input watts over a week
snmpwalk -v3 -l authPriv -u monitor -a SHA -A '***' -x AES -X '***'   ups.example.com 1.3.6.1.2.1.33.1.4.4.1.5

# per-port PHY behaviour on a Linux host
ethtool --show-eee eth0
ethtool -S eth0 | grep -i -E 'eee|lpi|wake'

If your PDU is monitored - APC network management cards expose power per outlet group, and the polling setup is the same as any other device as described in this UPS SNMP monitoring guide - you can get a defensible before/after number rather than an estimate.

Pairing With PoE Power Saving

On PoE access ports there is a second, larger lever: 802.3bt negotiation lets the switch allocate power based on the class the device actually presents. Standardising on the right class, and understanding the difference between the PSE allocation and the PD's guaranteed input, often saves more than EEE does. The class-by-class numbers and the cable-loss budget are broken down in this PoE standards comparison.

Summary

Enable EEE by default on ordinary copper access ports, document the exceptions, and never enable Ultra Low Power Mode on a link that has to react to link events quickly. Then measure at the PDU so the saving is a number you can defend, not a claim from a datasheet.

原文链接：https://superops.com/tech-hub/what-is-energy-efficient-ethernet