Ning made some interesting observations in his original posts about Cisco, QNX and VxWorks and the future for network equipment in general. Unfortunately, his rebuttal to my comments simply missed the main points, or he misinterpreted what I was trying to say. I apologize if my quick response caused the latter.
He made the following key observation in his original post, with the most important sentence in bold font.
Therefore, the modern networking equipment demands more of advanced features commonly found on server or desktop OSes, such as Unix, and less of absolute real-time response of traditional RTOSes, such as VxWorks. In fact, embedded OS is more appropriate here than real-time OS.
QNX has microkernel, process model, dynamically loadable modules and device drivers, transparent distributed processing and many other features that are lacking in VxWorks. For years, some people have been arguing against putting such features in RTOS for performance and footprint reasons. As hardware gets more powerful and software gets more complex, the arguments are no longer valid, especially for the networking equipment.
Think about what he wrote. Read it again, and think about it. First, Ning, not me, is saying that for years some people have been arguing against the feature-bloat he lists are in QNX. He then says, effectively, that Moore's Law no longer makes these arguments valid, especially for networking equipment.
This is the point of our debate. This is the only point of our debate. If Ning's assertion is true, then general-purpose OSs, like what you find on desktops and servers, will take over network equipment. I made the point that Bill Gates made this bet early on with Windows -- and won big time. Bill won because Moore's Law advanced hardware on the desktop faster than even his bloated software could absorb. (I did not say anything about evil monopolies. Indeed, this positioning by Gates is one of about two or three moves that led Microsoft's PC monopoly.)
Let me say it again. Ning made an important assertion that implies something critical for the future of WIND in network equipment and by extension to most other market segments. The issue is not whether QNX is the Holy Grail of OSs that can handle a desktop while also shrinking to a footprint that embeds in any device and runs faster and better than any other OS or RTOS. Not that he isn't allowed to make that claim also, but that's a detail that only becomes interesting if QNX gets on the radar screen. Certainly it is a detail that pales when compared to his main claim.
I agree with Nang that network equipment software is on a path that necessitates rapid increases in complexity (his latest rebuttal mysteriously implied I thought otherwise), actually for all the reasons he gave, and I am sure a bunch more. This has been a theme of mine from the start of this thread, if only because Intelligent I/O (I2O) embodies this theme and plays in this space. I have been particularly expressive about this theme lately when I talk of verticals, moving up the value-chain, etc.
The problem is the easy solution, of countering complexity with complexity, that worked well with servers and desktop OSs, faces an insuperable obstacle with network equipment. The lifeblood of network equipment, the bandwidth of information flow, promises forever to outpace Moore's Law. Think what this means. Even as microprocessors and related silicon double in most metrics every 18 months, they will forever fall farther and farther behind in their ability to keep up with much faster growth in bandwidth.
At first, network equipment makers resorted to ASICs to power past the capabilities of traditional software-based processors, with slower processors maintaining routing tables and other housekeeping tasks. But ASICs' inflexibility has given rise to a new breed of intelligent network processors that most likely embody the future. These processors have software everywhere. While a traditional RTOS does not run on the packet processors, there is every intention that the core processor will be deeply involved, through processor intercommunication, in decisions made at the packet level. This means, that the core simply must be slim and swift, reliable, and all the things that RTOSs do better than any other kind of operating system, be it a displatching loop or a feature-rich, general-purpose OS.
Moore's Law's loosing battle against the Law of the Telecosm guarantees that speed will be a primary consideration in software solutions. Coping with a zillion connected devices guarantees that complexity will rise meteorically, requiring software solutions. To me, the only certainty about how this will all play out is that (1) simple home-grown OSs are inadequate to handle required software complexity and (2) fully-featured general-purpose OSs are too cumbersome, inefficient and cannot adapt easily and reliably to exotic hardware. This leaves only the RTOS-level of solution, like VxWorks, that makes any sense for the network equipment space.
I think the above fairly restates the major differences in opinion between Ning and me. Add any other sub-issues, and the everything gets obfuscated. For example, he summarized what he thought I wrote and then responded point by point, to wit:
1. QNX is a bloated general-purpose OS.
I don't believe I said that. I used the word bloated in association with the genre of server OSs, which is true by definition or else they wouldn?t be general-purpose. Nevertheless, in commenting about what I didn't say, he claimed QNX has server OS features with the minimal footprint of a traditional RTOS.
If QNX has perfect design and all needed features, and is as small and efficient as, say VxWorks, then why bother to justify the use of server OSs on the basis of Moore's Law-type hardware enhancements? Why not just say QNX does more, faster with the same resources? That's a different argument, and one I'm not equipped to tackle. I don?t believe I have ever claimed VxWorks is superior to any other OS, but only that there is plenty of evidence that it is at least adequate for a large range of embedded applications.
Nevertheless, let me show you how topsy-turvy the embedded world can be when it comes to OS features. I believe an important feature of QNX, that made it attractive to Cisco to serve as the foundation of a redo of IOS, was memory protection. Some might argue that if VxWorks had "protection" over the last couple of years, there wouldn't be any other RTOS left standing, including QNX.
Now imagine yourself making a network device, and needing to extract every ounce of speed out of the processor. QNX's memory protection costs lost processor cycles. Can you afford to loose those cycles and be competitive? On the other hand, can you afford not to check each memory access for a violation?
Of course the answer depends on how often you expect the device to encounter a memory protection violation. General-purpose computers encounter them all the time, but even there the OS may not bother checking if it can be assumed the end-user is forgiving (sound familiar?). Fixed-function devices like network equipment (for these devices, complication arises at the design and development stage, not when deployed) run the same, fixed programs repeatedly. It is likely that many such devices NEVER will suffer memory violations that need protection, or almost never. Depending on expected frequency of encountering memory violations and the cost of a non-caught violation, the best solution for most network equipment may be to skip protection.
In other words, few, if any, OS features should be considered desirable automatically. Particularly, in the Network Equipment space, tradeoffs will always be considered when there is a price to paid in reduced speed.
2) Increasing complexity of network software is, like that of PC software, the evil design of a monopoly.
I never said anything remotely like this. Not even sort of remotely. I believe strongly in the increasing complexity of network software.
I referred to Bill Gates, not as an evil monopolist, but as a savvy guy who knew PC hardware would grow into his feature-rich software. It was a real-world example of Ning's original thesis.
3) The complexity is the main driver for more power hardware (Allen used the PC analogy to make his point that unnecessarily complex software is creating demand for more powerful hardware).
This was not the point of the PC analogy, as explained above. If that wasn't clear to Ning, I apologize for sloppy writing.
4) Intel's IXP (which is part of their IXA), IBM's Rainier and other network processors will once again put RTOS in the path of packet processing and forwarding.
I didn't say this. I said it was inconceivable to me that the OS on the slower-running processor core would not be sensitive to timing. There is a lot of coordinated stuff potentially going on at the processor level, but packet processing obviously isn't one them. I say "potentially" because who knows how those packet engines will be used in future products, and how much sharing of information might be required between the packet engines and the core processor? It seems essential to me that the core processor OS be an RTOS. Intel and IBM seemed to agree.
5) The old RTOS without advanced features can handle all of the foreseeable needs of networking.
Actually, there is constant feature growth in VxWorks, feature growth that is market driven. HA and protection are two recent examples. Otherwise, needed functionality is layered on through middleware or applications.
I don't know where the perfect balance is between bloat and all the other characteristics needed in the network equipment space. I know it is not a non-RTOS, general-purpose OS, and I know it is not a dispatching loop.
6) Cisco made an mistake selecting QNX while most companies it acquired are using older generation RTOS.
Yes, I think I said or implied I am not impressed with the QNC decision. Moreover, I think they are making an even bigger mistake in how they are integrating QNX into IOS. I don't care how big you are, there is no justification anymore for mucking around with a home-grown OS and tools -- unless that is your core business. Starting with a commercial QNX, no matter how good it is, doesn't change anything one iota.
7) To be Microsoft or not to be Microsoft.
Becoming the Microsoft of the post-desktop era is the only acceptable outcome for WIND, and it is attainable. This is a simple consequence economics in a competitive market with a zero marginal cost of production. There is no in between; its all or nothing.
In a space still dominated by home-grown OSs, it is easy to see why agnosticism toward commercial OSs will die hard, if ever. This is why the only thing about this debate is Ning's original premise that network equipment will gravitate toward the need for general-purpose OSs. If Ning's sub-premise that, by the way, QNX is that and also the greatest RTOS ever, then WIND will either buy QNX or crush it.
Assuming that VxWorks is essentially adequate for most the targeted market, especially network equipment, then how does WIND expect to dominate in an atmosphere of OS agnosticism? Certainly not by trying to convince old and new customers alike that VxWorks is sufficiently better than a long list of competitors to warrant, not only selection, but relatively high royalties.
The answer is that agnosticism works both ways. If WIND is bundled with a total solution, like Digita for cameras or Enhanced TV from Liberate, the manufacturer could not care less what OS underlies the solution. The manufacturer is simply trying to get to market fast with a robust product.
Semiconductor manufacturers are beginning to understand the importance of total solutions to them, and that their clients ARE OS-agnostic. They no longer care a whet about pretending to support all the world's RTOSs. All they care about is making sure that the industry standard total solution runs on their chips. Therefore, it takes them about 20 seconds to realize that they need to line up with the best soup-to-nuts supplier acceptable to their customer base. Nine out ten times this will be the industry leader, WIND.
As this process picks up steam, network effects begin to kick in, first virtually, as in available experienced engineers, and then real, as when total solutions begin to cross-pollinate. Network effects, which augments already awesome economies of scale, forms the final, nearly insuperable barrier to competition. As generally recognized by economists studying the New Economy, the resulting monopoly will exist at least until the next discontinuity in the underlying paradigm, or until the company, with a recognized monopoly, oversteps acceptable behavior by monopolizing a market unfairly and becomes the target of antitrust action.
I think this process for WIND is well underway, and probably already is unstoppable. What is particularly interesting in the context of this debate is that the industry's OS-agnosticism feeds, not hinders, the process. What is even more interesting, if you think really deeply about it, is that Microsoft became the world's most valuable company by dominating a paradigm that will be dwarfed by the size and breath of the next computing paradigm. In this sense, I agree that WIND will not be the next Microsoft. WIND is headed for a far more exalted accomplishment.
Allen |
