Duplex conflicts simply will not go away. After a decade of plaguing IP networks, duplex conflicts still seem to be the single worst source of performance degradation. Part of the 10 Mbps Ethernet legacy, this problem is trivial to fix but ridiculously hard to identify and localize. And it recurs frequently, as interfaces go up and down over time and network hosts are updated and changed.
It is a rare network administrator (or one new to Ethernet technologies) who does not have firsthand experience of diagnosing duplex conflict (otherwise known as "duplex mismatch"). It can have devastating effects on all sorts of applications. It is highly transient in its symptoms, sometimes showing no loss when using ping (see
What Ping Doesn't Tell You) and at other times causing upward of 60% loss on a 100 Mbps path. It is typically triggered by heavy data transfers and can affect VoIP even when VLANs and QoS are employed.
Down-level NIC drivers are the most common form of performance degradation (see Numbers lie: Your NIC could be killing your network performance), but duplex issues are nearly as pervasive and typically have a much more dramatic impact on end-user experience. In 2001, a study at NASA concluded that more than 50% of trouble tickets were attributable to duplex conflicts. Internet2 ranks duplex issues as one of the top three most common problems (the others being incorrectly set TCP buffers and NIC drivers). In a 2005 real-world study based on 20,000 support calls received by Veritas (now Symantec) for network-dependent products such as NetBackup, nearly 38% of the calls were attributed to network issues. Of those involving the network, 20% were caused by poorly performing NIC drivers, while nearly 6% were identified as caused by duplex conflicts.
Why is it such a problem, particularly when auto-negotiation should be the solution?
Most Ethernet NICs from the last 10 years have come equipped with a mix of speeds and duplex modes that are intended to ensure compatibility among devices. The most common speeds are 10 Mbps (Ethernet/802.3i/10BASE-T) and 100 Mbps (Fast Ethernet/802.3u/100BASE-TX), with 1000 Mbps (Gigabit/802.3ab/1000BASE-T) becoming more typical. Duplex modes include half-duplex and full-duplex. Half-duplex supports communications among devices sharing the same media for both transmission and reception (e.g., one pair of copper wires); full-duplex assumes separate connections for each (e.g., two pairs of wires).
The simplest case of duplex conflict is where two interfaces -- for example, a port on a switch and an NIC on a workstation -- are manually set to different duplexes (but the same speed). (In the figure, the switch is set to half duplex and the workstation to full duplex.) While the switch uses collision detection (CSMA/CD) to avoid transmitting when the workstation is sending, the workstation sends whenever necessary, without regard to the switch. Consequently, packets sent by the workstation may collide with packets from the switch -- the switch attempts to re-transmit those it sent when collisions occurred; the workstation does not re-transmit.
To the application user, the network feels "slow," even though the application will typically start without any difficulty. The classic symptom is an FTP client that establishes a connection and begins transfer but then crawls along, taking many minutes to transfer a relatively small file. In this case, TCP is constantly backing down its transfer rate as it incorrectly interprets packets lost to collisions as congestion.
At the network level, duplex conflicts can be manifested as severe packet loss, particularly as the rate of two-way traffic increases. Ping, usually sending one packet at a time, may not see any loss at all, unless some other application is using the same link. On the switch, the interface will show high collision counts (of the type "late collisions," resulting in frame corruption), and the interface on the workstation will record high numbers of CRC errors.
Properly configured, these same interfaces will function at the selected speed without any loss at all. Manually setting interfaces can be problematic for naïve users, as well as unmanageable in enterprises where there are thousands of interfaces to configure. It is therefore preferable to have the interfaces discover the best choice of speed and duplex for themselves.
When two Ethernet interfaces configured for auto-negotiation initially become active (and at some other times), they will attempt to negotiate with each other to establish a common configuration. Auto-negotiation and auto-sensing provide mechanisms by which two interfaces can properly select the optimal configuration under a variety of circumstances, including when one of the interfaces does not know how to auto-negotiate.
Automated speed selection, or auto-sensing, tends to be reliable, resulting in no connection at all when it fails. Thus, speed mismatch happens rarely, and when it does happen, it is relatively obvious. Auto-negotiation for duplex selection, on the other hand, has not been nearly as reliable, resulting in connections that are functional but highly degraded under certain conditions. In fact, it is not usually auto-negotiation itself that is at fault; rather, one interface has been set for auto-negotiation and the other has been configured for a specific speed/duplex setting, effectively disabling auto-negotiation.
In particular, the auto-negotiation protocol requires that the interface fall back to half-duplex mode when it is not successful in negotiation. Thus, if one side is set for 100 Mbps full duplex and the other side for auto-negotiation, a duplex conflict is almost certain. A classic blunder occurs when the naïve user installs a hub between a workstation and switch in order to add a printer or other Ethernet device. The IT administrator may have opted to use a fixed configuration (such as 100 Mbps full-duplex) between the workstation and switch to avoid problems with auto-negotiation. Consequently, as all hubs are by nature half-duplex, the user instantly causes a duplex conflict.
In part 2 of our series on duplex mismatch, the implications of duplex conflict for Gigabit and QoS, as well as auto-negotiation best practices is explored.
Chief Scientist for Apparent Networks, Loki Jorgenson, PhD, has been active in computation, physics and mathematics, scientific visualization, and simulation for over 18 years. Trained in computational physics at Queen's and McGill universities, he has published in areas as diverse as philosophy, graphics, educational technologies, statistical mechanics, logic and number theory. Also, he acts as Adjunct Professor of Mathematics at Simon Fraser University where he co-founded the Center for Experimental and Constructive Mathematics (CECM). He has headed research in numerous academic projects from high-performance computing to digital publishing, working closely with private sector partners and government. At Apparent Networks Inc., Jorgenson leads network research in high performance, wireless, VoIP and other application performance, typically through practical collaboration with academic organizations and other thought leaders such as BCnet, Texas A&M, CANARIE, and Internet2. www.apparentnetworks.com
This was first published in July 2006