Thursday, 20 August 2015 04:23

MEF Addresses Potential TCP Performance Issues

Written by  Joan Engebretson
Rate this item
(2 votes)

MEF Addresses Potential TCP Performance Issues

MEF researchers have devised a solution to a problem that has plagued Carrier Ethernet for years – TCP performance that suffers when a customer has a Carrier Ethernet virtual circuit (EVC) provisioned over a higher-speed physical link. Ethernet Academy talked and exchanged emails recently with Stephen Haddock about the problem, the solution and the next steps in implementing the solution. In addition to being the former CTO of Extreme Networks, Haddock is principal of Stephen Haddock Consulting and a contractor to the MEF.

The Problem

The TCP problem pertains to an interaction that occurs between TCP’s adaptive congestion control/ congestion avoidance and the Carrier Ethernet policer. As we detailed in a previous Ethernet Academy post, the role of the policer essentially is to slow down traffic if, for example, a 100 Mbps physical Carrier Ethernet connection is provisioned only for 10 Mbps service.

What happens in this example is that TCP continually increases the window until it gets to 100 Mbps, which is faster than the committed information rate (CIR). At that point the committed burst size (CBS) bucket is completely depleted, the effective bandwidth of the circuit goes to zero and a timeout is created. 

The figure below shows simulation results for a TCP transmitter sending a stream of data to a TCP receiver through an EVC with a UNI speed of 100 Mbps but a CIR of 10 Mbps and a CBS of 10 segments. As Haddock explained, “in this simulation, one segment consists of 1460 bytes of TCP data which becomes a 1522-byte packet on the Ethernet link,” yielding a CBS of 15.2 kilobytes. 

 

"The figure shows the cumulative number of TCP segments received at the receiver over time, so the slope of the curves is the TCP throughput," Haddock explained.

The orange curve depicts the problem.  Ideally the throughput would equal the CIR. But that doesn't happen.

"TCP attempts to ramp up its transmission rate as high as the UNI will allow, but after an initial burst the bandwidth profile policer starts dropping packets in excess of CIR which causes TCP to timeout," said Haddock. "After the timeout TCP tries again and the process repeats. The result is an average throughput well below CIR."

Although this issue could be addressed by increasing the CBS substantially (by a factor of 20 in this case), that solution creates other problems as multiple users could be involved, potentially requiring buffer sizes larger than equipment can support.

The Solution

The best solution, according to MEF research, is to shape the traffic at the customer premises before it enters the EVC. 

Perhaps surprisingly, Haddock said "the details of the shaper implementation don't matter much." The important thing, he said, is that "you have a buffer to hold incoming data that exceeds the CIR."

The blue line in the figure above shows the identical situation depicted with the orange curve, except that a shaper is inserted before the bandwidth profile. As Haddock explained, "When TCP attempts to ramp up its transmission, the shaper buffers the traffic that exceeds CIR, rather than discarding it as the bandwidth profile would."

TCP, he said, "reacts more elegantly to packets delayed by the shaper than to packets discarded by the bandwidth profile policer," yielding an average TCP throughput that equals the CIR.

Next Steps

Depending on the situation, the equipment at the customer premises may be owned by the end user or by the carrier – and if it's the former, Haddock expressed concern that some customers might balk at the need to do traffic shaping.

Carriers, he said, will need to take care in explaining the issue and advising customers that they "will get much better performance if they do the shaping."

To help carriers with this task, the MEF is in the process of creating an appendix to the upcoming MEF 23.2 standard that explains the problem and the best way to set the CPE to address it. The title of the document will be "MEF Technical Specification: Carrier Ethernet Class of Service – Phase 3." Additionally the section of the standard that addresses burst size will be updated to reflect the MEF's findings.

Read 6683 times Last modified on Thursday, 20 August 2015 05:28
Login to post comments