Summary

While Emulex reported better than expected EPS results, the tone of its earnings call was not upbeat. Weak demand in Unix markets, combined with revenue of $102.2M, which was at the lower end of expectations, and weakness in HBA markets has Wall Street analysts concerned. As is often the case in Emulex’ core markets, gross margins were strong with non-GAAP gross margins at 68% but guidance suggests limited upside in the gross margins line.

As well, the company is guiding revenue of $100M-$10M sequentially for Q4 suggesting that this is a second half story with a hockey stick-like demand curve attributable to new CNA design wins and a rebound in Unix markets attributable to Unix server refreshes from server lines such as IBM’s System p.

Wall Street is also concerned about the sustainability of Emulex’ embedded business which has been growing nicely over the past three quarters. Questions on the call clearly indicated consternation over inventory levels which rose 54% sequentially. While management tried to convince analysts that inventory levels are returning to normal from abnormally good levels, it appears that the Unix shortfall caught Emulex off guard.

Of equal concern is that when management was asked how much exposure it has to the Unix business it essentially said “we don’t know—but it’s meaningful.”

Positives

Emulex remains the only other major player, along with QLogic in the storage HBA business. The challenges associated with developing hardened storage stacks translate into extremely good gross margins. While Seagate struggles to achieve 30% gross margins on incredibly complex products, Emulex and QLogic consistently hit mid-60’s on the gross margin line. There is little doubt that Emulex enjoys a position in a duopoly that isn’t likely to change in the near term.

Expense control. Emulex is managing operating expenses in line with soft expectations for the June quarter and continued financial discipline will ensure Emulex’ viability indefinitely.

CNA design wins. Emulex claims 15 tier 1 CNA design wins in storage and Ethernet 10Gb NICs. Of course there aren’t 15 tier 1 OEMs but Emulex is counting multiple product line wins within tier 1 OEMs. Emulex got off to a slow start in design wins relative to QLogic but as expected is picking up design wins as most OEMs require a second source in this market.

Squinting Through the Benchmarketing Hype

One of the things Emulex CEO Jim McCluney touted on the earnings call was Emulex’ 1M IOPs performance figure from benchmarks that are available on the Emulex Web site in the form of a Report Card (Google search warning). I’ve been looking through the Emulex figures for the past several weeks and also asked QLogic to provide me with any benchmark data it has. The 1M IOPs number is a marketing number that has no bearing on the real world and I am compelled to call this out.

I’m still sifting through the numbers but I noticed several things in the data:

• The Emulex and QLogic benchmarks appear to be similar but there are discrepancies in the two data sets which I need to reconcile.
• For example, both sets of data show Emulex approaching 1M IOPs at .5KB block sizes (yeah, 0.5KB) however it’s unclear what the workload is in the Emulex data set. For the QLogic benchmarks Emulex runs at about 841,000 IOPs at 0.5KB in a sequential read workload while the Emulex report card shows 991,000 IOPs at 0.5KB – (yes I know…0.5KB is silly—no one runs workloads at ½ KB). I’m not suggesting any subterfuge by Emulex, it’s just the charts on the report card are not clear.
• When pressed, Emulex will freely admit that a 0.5KB block size is not a common situation and that 4K and 8K are more typical of real world workloads.
• I also noticed that the QLogic benchmarks included an IOMeter output that measures CPU utilization. At .5KB and 841,000 IOPs the CPU utilization of the 2.8 GHz – Intel Nehalem Dual Socket Quad Core being tested was 81% – yup; 81% CPU utilization.

What does this mean? It means not only is .5KB not a real world workload but even if it were, no one would be able to run any kind of demanding application because the Emulex CNA consumes so much CPU horsepower at smaller block size levels. There is always an inverse relationship between block size and CPU utilization but the figures I’m looking at are absurd…until of course you get to 4KB block sizes and above. In these instances Emulex’ CPU utilization, while still higher than QLogic’s is not objectionable.

Why is Emulex’ CPU Utilization so High?

I don’t know the answer to this but I suspect it relates to the choice of using a ServerEngines chip set. As a time-to-market play, Emulex outsourced its CNA chip design to ServerEngines which has a fair amount of processing power on its chip. However it’s possible that Emulex has offloaded some of the processing function from firmware to its driver to optimize performance. In doing so, CPU utilization would probably escalate quite dramatically, especially at the smaller block sizes that Emulex likes to tout in its performance advantage.

It would be fairly easy to test this theory as the Emulex Linux driver is publicly viewable. I personally wouldn’t know what to look for but a technical whiz could stare at the code and see if basic bookkeeping functions – like frame header info and assembly/disassembly processing – or other error handling and vector IO techniques are being pushed to the driver.

Does this Matter?

Really no. But when a vendor starts touting something as absurd as 1M IOPs as a 2X competitive advantage it’s important to understand what is reality and what is hype. As we’ve previously published in our CNA Deep Dive performance in these markets is basically ‘table stakes’ with Emulex and QLogic about comparable. What’s of greater concern is the attention being paid to such meaningless figures as 1M IOPs at 0.5KB block sizes with unsustainable CPU utilization. There is no value in these metrics to IT practitioners; none.

What would be of value is to look at application performance. Specifically, as CNAs are used to virtualize IO, figures that give users confidence that virtualized applications will run at acceptable levels would go much further to advancing technology adoption than meaningless data.

I’m calling on both Emulex and QLogic to cut out the “Urinary Olympics” and start developing metrics that matter. That way CIOs can be certain that storage stacks will support a new breed of virtualized applications, end-to-end, and not get sucked into worthless marketing hype.