Archive for November, 2006

This story appeared in Network World here.

“We are used to breaking products in lab testing, but this time we broke 802.11 itself. Our tests uncovered a design flaw in the Wi-Fi protocol that affects performance testing, not just for current 802.11a/g products, but possibly in upcoming 802.11n gear as well. As a result of our tests, an IEEE committee heard a proposal to recognize and fix the design flaw.

It’s a common misperception that Wi-Fi is an inherently “lossy” medium. Wi-Fi is highly vulnerable to signal errors, but it compensates with built-in error checking and retransmission mechanisms. Even a huge error rate (say, 10% of all packets, the maximum allowed in 802.11) should still result in zero loss, because packet errors are retransmitted.
That’s the theory. In practice, we found a deficiency in an 802.11 packet header that can lead to packet loss.

The physical layer convergence procedure (PLCP) header carries key information about each packet, such as its length and transmission rate. While the rest of an 802.11 packet has excellent error protection because of a 32-bit CRC field, the PLCP header has only a single bit for error checking, and that is nowhere near enough to protect against corruption.

Weak PLCP error checking can fool an 802.11 receiver into believing that it never received packets, even after a transmitter goes through multiple retry attempts.

For example, suppose an 802.11g transmitter sends a 100-byte packet at 54Mbps, and that channel noise corrupts the PLCP header. The corrupted header can convey bogus values, such as telling the receiver the packet is 4,095 bytes long and is being sent at 6Mbps.

An uncorrupted packet would take just 36 microsec to transmit, but in this case the corrupted PLCP header will cause the receiver to keep listening for the packet for 5,484 microsec. The receiver is literally off the air for that long period, causing it to miss multiple retry attempts and give up on the packet as lost.

This perceived loss makes it harder to get an accurate read on device performance. It’s standard practice in throughput and latency tests to tolerate zero dropped packets. Because weak error checking in the PLCP header introduces packet loss, lower throughput rates are a likely result.

Weak error handling also can affect roaming tests. If a receiver misses an Extensible Authentication Protocol handshake packet during a roaming event, it can take 30 seconds before the RADIUS handshake begins again. We saw some 30-second roaming times in our tests because of this issue.

The probability of PLCP corruption with short packets and high rates is around one in 1,000. Because performance tests inevitably involve far more than 1,000 packets, results easily can be skewed downward by corrupted PLCP headers.

We compensated for this issue by setting an acceptable loss threshold of 0.1% in our throughput tests. We’re not crazy about allowing loss in throughput tests. It’s a violation of RFCs 1242 and 2544, and it’s a common dodge used by vendors of poorly performing products.

In the absence of a better solution, however, we opted for the acceptable packet loss to mask any instances of PLCP header corruption.

While there’s no chance of changing the 802.11a/g standard because of backward compatibility issues, it’s not too late for the IEEE to correct this design flaw in the forthcoming 802.11n standard. The IEEE recently heard a proposal to address the issue uncovered in these tests.

At the IEEE 802.11 meeting in Melbourne in September, test equipment maker VeriWave and Wi-Fi switch vendor Aruba Wireless Networks jointly outlined the problem and described its implications for performance testing. The current 802.11n draft uses the same PLCP structure as 802.11a/g when operating in mixed mode (with 802.11b/g clients present). The presenters recommended stronger error protection for 802.11n mixed-mode operation, as well as warning language in the forthcoming 802.11.2 standards to alert testers to the design flaw.

An obvious question is why this phenomenon hasn’t been seen before. The answer has much to do with the relatively crude state of Wi-Fi performance testing. Besides the perception that “it’s just wireless,” previous tests haven’t offered enough packets at high enough rates, and closely analyzed the results, to understand that the current 802.11 protocol is itself an impediment to improved performance.”

Comments No Comments »

As detailed here, Sun Microsystems plans to introduce a new feature in early January to let its UltraSparc T1-based servers run multiple operating systems simultaneously, partly matching abilities of rival machines from IBM and Hewlett-Packard.

I1_ultrasparc_ivplus_lg.jpg

The technology, called logical domains, or LDoms, is the latest move in a race to make servers as flexible and efficient as possible through a general technology called virtualization. Customers using the technology will require the 11/06 update to Solaris 10, due in November, and either new hardware due in January or a firmware update to existing Sun Fire T1000 and T2000 servers, said Pradeep Parmar, product manager of business strategy for Sun’s server group.

For Sun, the LDom approach adds a third, intermediate option when it comes to sharing multiple tasks on a single machine.

Comments No Comments »

Cisco Systems has agreed to buy Greenfield Networks in a move designed to expand its Ethernet lineup, the companies said earlier this week.
Greenfield, a privately held Sunnyvale, Calif.-based company with 60 employees, develops semiconductors designed to improve Ethernet packet processing for the so-called metro Ethernet market. Telecommunications carriers use the technology to string together a number of corporate networks within a metropolitan-area network, allowing these companies to use VoIP and data over their networks.

“By integrating Greenfield Networks technology with Cisco’s family of metro Ethernet switches, we will be able to improve the time to market of new carrier-class features to our service provider partners,” Kathy Hill, senior vice president of Cisco’s Ethernet and Wireless Technology Group, said in a statement.

Greenfield, which also has offices in Bangalore, India, was founded in 2000. Cisco plans to fold Greenfield’s product portfolio and employees into Hill’s group, once the deal closes during Cisco’s current fiscal quarter, which ends in late January. Financial terms of the deal were not disclosed.

The networking giant has been building up its metro Ethernet offerings over a number of years, as the concept of metro Ethernet has become mainstream.

Comments No Comments »