After reading your second letter regarding Linux on the mainframe - where you attempted to defend Sun’s Linux position shortly before the departure of Ed Zander and Stephen deWitt from Sun’s executive staff – it is now obvious that your strategy is to undermine the credibility of Linux as a serious OS and position Linux solely on the low-end and your Solaris on the high-end. I understand that Sun is driven to do this because poor long-term architecture strategy that has resulted in a loss of market demand for Sparc/Solaris and serious financial problems for Sun. There is a negative feeling in the Linux community with regard to Sun’s tactics and many Linux developers have asked me to respond to the technical inaccuracies in your most recent letter. The Linux community understands your real attack is against Linux on high-end servers, in general. Sun is protecting its own self-interest and this has never been more apparent.
Words from the Net
”From a Linux advocates point of view, there isn't much difference between Sun and Microsoft. Don't be fooled by the saying ‘My enemy's enemy is my friend’, because it doesn't apply here.”
Community Posting on Slashdot.orgt
I responded to your original letter - not because of blind love for Linux and/or naïve misunderstandings of the IBM zSeries strategy – but because my work involves research on open standards “mainframes” based on Intel processors and dynamic partitioning technology with fair-sharing of all system-resources – the work inspired by the actual mainframes.
SWsoft does make money from Linux development, but we also freely give back to the Linux community by supporting engineers like Alexey Kuznetsov, Chief Software Engineer for SWsoft famous for freely contributing his work to improve the Linux kernel networking components.
We work with the Linux community, including the major distributions, to advance the interests of Linux in mission critical situations in the enterprise with both carrier-grade Virtuozzo-powered Linux and in large-scale Internet hosting deployments with our HSPcomplete solution.
|Although our development is mostly focused on Linux on Intel, moving toward Intel-based open standard mainframes, rather than on mainframes, we have been ordained by the Linux community to defend any attack on Linux in the enterprise. I am financially motivated to see Linux succeed; however, my intentions are honest and my company is unyielding in its support of Linux. Kuznetsov, leader of the Linux kernel networking team, adds, “We are working hard to technologically advance Linux in a way that pertains to the wishes of the community and not independent, commercial interests.”||
Words from the Net
”sun does not understand linux. it is only trying to protect it’s own interests, which is polar to the spirit of the open source movement. it’s more like ‘what can sun do to us’ than ‘for’ us…”
– Newsforge posting
Sun’s Weak Position Had Lead to Serious Financial Problems
Server markets can be categorized in many ways; however, there are four segments that are obviously the most relevant to Sun. Interestingly enough, these same four market segments are also the segments most important to Linux:
For years, Sun was content to dominate the market for general purpose open systems for mid-to-large enterprises. Unfortunately for Sun, Linux emerged as a formidable opponent. Linux on Intel servers are now winning in the mid-enterprise market for general purpose “Open Systems.” Mainframes have long dominated the enterprise market and Linux on the mainframe suddenly turned the previously closed mainframe into the perfect “Open System” for large enterprises. Customers wanting immediate access to Linux on enterprise hardware did not have to wait the three years technologists predicted it would take Intel-based servers to catch up to the mainframes high availability features. These large enterprise customers rapidly rushed to embrace Linux on the mainframe.
Sun’s position will continue to worsen as multi-processor Intel servers improve and new dynamic partitioning technologies like carrier-grade Virtuozzo for Linux on Intel enable Linux to beat Sparc/Solaris in both the mid enterprise and large enterprise markets. Due to its open nature, Linux already dominates in the hosting market and with rock-solid service provider solutions like HSPcomplete for Linux that enable much better cost management and hundreds of CPUs than a comparable Solaris solution. Within the next five years, Linux on Intel servers will compete against Linux on the mainframe in large enterprise, further accelerating the improvement of Linux in large enterprise markets.
Solutions like HSPcomplete for service providers are being developed for Linux instead of Solaris because Linux is a much more open system than Solaris. In fact, despite Sun’s claims, Solaris and Sparc have actually become closed systems, especially when compared to Linux on Intel. Independent software developers such as SWsoft have the right and ability to modify the source code for Linux systems. Linux is a much more attractive development platform and there is a greater number of Linux developers, which ultimately means the creation of more higher-quality applications.
My view of Solaris as a closed system for independent software vendors is based on actual experience. I tried to obtain the source code for Solaris in order to evaluate the cost of porting carrier-grade Virtuozzo to Solaris, but I was not able to get proper access to the required technical information or source code for the needed parts of Solaris. Out of frustration, I decided to put development of HSPcomplete and Virtuozzo for Solaris on hold and decided to focus on Linux. After six months, we again looked at the feasibility of a Virtuozzo port to a closed system. Like many other companies and due to the market share of Windows in the closed systems market, we decided to put resources on Virtuozzo on Windows instead of Solaris. We also looked at improving Sun’s Cobalt line of servers with dynamic partitioning technologies. Unfortunately, due to the closed nature of Sun’s version of Linux, this was also put on hold. I invite Sun to open their technologies more and help other ISVs. If Sun were to open their technologies more, we would definitely revisit developing solutions for Sun’s platforms.
|Words from the Net ""Set still," said Shark. "You ain't goin' to hit no breeze, Bob. I hate to tell you, but there ain't any chance for but one of us. Bolivar, he's plenty tired, and he can't carry double."" O'Henry, The Roads we Take||
Many developers and customers classify general purpose servers into two categories:
|Although Sun now has a Linux offering, you are likely to lose to IBM, whose success is driven by IBM Global Services, the world’s most expansive technical services organization. Without a strong service organization, Sun finds itself unable to profitably embrace high-end Linux servers. Sun also has no clear advantage over companies like Dell for entry-level Linux servers. Without a way to make a profit, Sun tries to weaken the business case for Linux by promoting the altruistic goals of the open source community. In reality, this is an attempt to confuse the business community and the truth is that Linux businesses are real and thus a real threat to Sun. IBM invests in Linux to make money. Red Hat, SuSE, Caldera, Turbolinux and SWsoft also invest in Linux to make money. Linux is not an altruistic community of hackers. Linux is appropriate for business. Unfortunately, it is just not appropriate for Sun’s business.||
Words from the Net
“When the business community first started talking about open source, the big question was how an open-source company could possibly turn a profit. Two years later, there are a number of profitable, successful open-source companies.”
ZDNet Tech Update, Michael Olson
Problems with Khan’s “Wolf in Sheep’s Clothing” letter Mr. Khan, your second letter once again includes a number of inaccuracies and exaggerations.
The most important flaw exists in the basis of your argument - “mainframe systems still only go up to 16 CPUs, when Sun made 64-way systems five years ago and has 106-way single-cabinet systems today.”
The "Amdahl law" states that any workflow can be broken into parallelalizeable and serial (non parallelalizeable) parts. This shows that CPU number is not as important as how efficiently the server is used. Suppose the system has N processors to execute the parallel operation:
Total execution time = parallel part/N + serial part
If there are the 106 CPUs that you mention Sun servers can support, the 'serial part' of the application, middleware, hardware involved in this workflow is significantly smaller than 1% of the total execution time. This basically means that machines need to be balanced and a high number of CPUs, in actuality, do not make it faster if the rest of the hardware, operating system, middleware and application is serialized to some extent.
You conveniently forgot to mention that applications do not scale linearly. Applications need to be specifically written using parallel programming techniques in order to run efficiently on Sun hardware with higher numbers of CPUs, which is expensive and risky. Computer programmers with experience in parallel processing techniques are rare and costly. If the program is incorrectly designed, the application may actually run slower with 50 CPUs, than with four. After my own painful experience optimizing applications for multiple-processor servers, I came to appreciate the ability of mainframes to transparently scale applications with zero programming work.
Most workflows cannot utilize 100s of CPUs. In order to actually use 100s of CPUs properly, a server needs to be able to handle multiple departmental workloads on multiple dynamic partitions with fair-sharing of system resources, including memory and CPU. Sun’s Dynamic System Domains lack the fair-sharing features found in LPAR and VM on IBM’s mainframes. These fair-sharing dynamic partitioning features are common on Linux on Intel with technologies such as carrier-grade Virtuozzo for Linux. Datacenters with large numbers of CPUs need VM or VE technologies to properly manage the security and reliability of their systems. The popularity of HSPcomplete at service providers proves that Virtuozzo Linux VEs are a rock-solid and highly flexible platforms for HSPs to deploy hundreds of Intel CPUs efficiently. Unfortunately for Sun, due to a lack of appropriate server technologies, the market for its high end servers is extremely limited. Due to the rapid development of mainframe-like features on Intel-based servers, mainframe customers that seek an alternate platform will turn to Linux VEs on large Intel servers instead of Solaris on Sparc.
Sparc/Solaris architecture is less efficient than mainframes for large servers. This is becoming more apparent as Sun builds larger servers. Unlike mainframes, which are designed for large numbers of processors, Sun’s architecture based on SMP results in a significant amount of CPU power to be taken up by memory I/O processing.
Moving applications to Linux on the mainframe can require zero programming effort. You attempt to scare and intimidate people new to open source and Linux by vague talks about the difficulties in porting new applications to Linux on the mainframe. However, you obscure the “open systems” nature of Linux and its acceptance of open standards such as Java, XML, HTML, and SQL. Most enterprise applications run on top of middleware such as WebSphere, DB2, Java application servers, and Oracle. Almost all of these applications were developed on top of middleware and will run without modification on the same middleware running on Linux. For example, an application written for WebSphere on Solaris will run on WebSphere for Linux without modification.
You, Mr. Khan, are technically ignorant of processor cache. It referring to the mainframe’s relatively small shared cache being antiquated with low bandwidth, it became apparent that you only understood Sun’s cache architecture and had no technical knowledge or engineering experience with the shared cache systems on mainframes. The cache design of Sun’s servers and the shared cache of mainframes are different and cannot be compared using your metric.
Shared cache systems can utilize all of the system’s bandwidth for commercial workloads. Dedicated cache systems have high bandwidth overhead leading to problems of cache coherency. Your claim that mainframes have reduced performance because of their “need to provide cache coherency across CPUs,” is likewise incorrect - the fundamental design of shared cache systems eliminates problems with cache coherency.
Your claim that the mainframe’s architecture is archaic because of the small cache size is also not true. The mainframe’s shared cache can be considerably smaller than the dedicated cache on Sun’s servers because of the lower cache missed rate of shared cache systems. The size of the mainframe’s processor cache is perfectly sized for the mainframe’s processors, as additional cache would have no performance value. If mainframes needed more processor cache, the cache size of the mainframe could easily be increased. Additional processor cache could be added quite cheaply. The reason that IBM doesn’t add more processor cache is because it is not needed.
You also did not mention that shared cache enables hardware to survive CPU failures, something that will crash a dedicated cache system, including Sun’s enterprise servers. This is one reason that mainframes are so much more reliable that Sun enterprise servers.
In addition to shared cache, IBM’s mainframes can use Parallel Sysplex clustering to further increase reliability and scalability, making the mainframe an unbeatable platform in the enterprise today.
”I knew the mainframe was years ahead of other servers in terms of reliability and scalability. It became clear that this world-class server needed a world-class operating system to run the new applications that make up the Internet. Linux was the obvious choice since it had already proven to be portable and ubiquitous.”
– Boas Betzler, IBM
Solaris is falling further behind Linux in the development of dynamic partitioning technologies with fair-sharing that help increase the availability and reliability of large servers. Thus, Linux on Intel, not Solaris on Sparc, is better positioned to challenge the mainframe in the future.
Mr. Khan, you forget to mention that Linux and z/OS are designed for concurrent use on the same servers. Many of your points on the weaknesses of Linux on the mainframe are invalid. You take a strategy of divide and conquer and point out weakness of Linux and point out weaknesses of the mainframe. For example, your recent article, “Recurring Mainframe Hosting Costs: How Sun Blue Them Away,” does not mention Linux on the mainframe at all. What you ignore in your letters and articles is that Linux and z/OS are normally run on the same mainframe at the same time. Linux cannot yet dominate all aspects of enterprise computing, but Linux partnered with z/OS creates a server that offers the best of both worlds and is unbeatable in the datacenter right now. In addition, Intel-based servers will improve over time and with Linux and Virtuozzo, will offer customers seeking high-end open systems another viable choice.
Linux helps to decrease TCO. Mr. Khan, you commented that for the price of one zSeries mainframe, we can buy 250 Sun Netra X1s, which is really an irrelevant comparison. Schools can by 50 bicycles for the cost of a school bus; however, we don’t have schools buying bicycles instead of buses because a value of the bus in terms of safety and convenience is more than the value of the 50 bicycles. The value of a zSeries is in its reliability and scalability, something not found in 250 Sun Netras, regardless of cost. In addition to the mainframe having higher value, the management and maintenance of 250 Netras is an enormous burden to an organization and increases the Total Cost of Ownership. Thus, a mainframe is not only more valuable, but also has a lower total cost of ownership than 250 Sun Netras.
Reality: Sun faces great technological challenges in today’s new computing world. Sun’s current markets are saturated and Sun is failing to penetrate new markets. Sun is betting on the high-end server market, but the Sparc/Solaris architecture lacks dynamic partitioning technology with key features such as fair-sharing of memory, fair sharing of CPUs and fair-sharing of memory I/O. Sun must resort to sophisticated marketing tactics to compete in the high-end server space. Due to the open nature of Linux, it is improving faster than Solaris at much lower costs. For example, Virtuozzo for Linux adds dynamic partitioning features to Linux on Intel for a cost of less than $300. Sun is only now starting development in this area. Solaris 9 will introduce software partitioning of resources, but it is not available yet and even when released will lack the resource control features suitable for mainframe-level production environments.
Despite Sun’s marketing efforts, the unfortunate reality for Sun is that Solaris/Sparc does not have the required features to compete with the mainframe on the high-end, is being killed by Linux and Windows on the low-end and, with new Linux OS Virtualization technologies, is being beaten by Linux in the medium to large enterprise mid-range.
“The partitioning and virtualization of the mainframe opens up a whole new world. With these capabilities, several hundred operating system images can run parallel to one another on the same box. IBM's experience in the server field has allowed us to consolidate a soccer field full of servers into logical images on a single box.”
Mr. Khan, you defended Sparc by pointing to the large number of developers working on this CPU and tried to show the weakness of IBM’s architecture by mentioning that IBM fired engineers from their CPU labs. You didn’t mention that the Linux community has more developers than Solaris and that the Linux community can’t be fired. When I notice that Sun does fire executives, I think of a famous saying, “Fish rots from the head.” Your ignorance of Linux community development is shown in your remark that the Solaris scheduler is better than the Linux scheduler. In your original letter, you reference work from Mike Kravetz of IBM. However, it appears that you didn’t even read Mr. Kravetz’s paper and the sources Mr. Kravetz references fully. If you had, you would have understood that the Linux scheduler is optimized for systems with less than 5 CPUs. For larger systems, the open modular design of the Linux kernel allows the default scheduler to be replaced with schedulers optimized for larger or smaller systems. I suggest you look at the paper that Kravetz references: “Loadable Scheduler Modules on Linux” by Scott Rhine of Hewlett Packard.
Two years ago, there were many schedulers available, such as Qlinux Fair Queuing, SGI CPU sets, Real-time extensions, User-based fair scheduler, and other developments like ShareII. There are also many commercial optimized Linux schedulers available today that are proven in the field. For example, the Virtuozzo scheduler that I helped develop for Linux has advanced scaling features and is in wide use in our HSPcomplete product that is widely deployed in Hosting Service Providers (HSPs) and other carrier-grade environments.
” Linux-based systems are the only viable Open Systems today. Sun’s systems can no longer be considered open,” said Serguei Beloussov, CEO, SWsoft. “Within 5 years, Linux systems will own all the markets that Sun is in now. Even more exciting for the enterprise customer is that within five to seven years, Linux with Virtuozzo-like technologies based on Intel architecture will provide leading industry-standard truly high-end and truly Open Systems with similar levels of availability, scalability, and reliability features as the IBM mainframe.”
The open source development model of Linux is a threat to Sun, a company that once dominated the market for general purpose open systems servers for mid-to-large enterprises. Unfortunately, due to poor long-term design decisions resulting in uncertain product positioning, Sun is experiencing financial problems. Sun is trying to move into the high-end server market, but lacks the required technology such as dynamic partitioning with fair-sharing. Linux on the mainframe and Linux on Intel with Virtuozzo is beating Sun in the high-end open systems markets. Thus, Sun is trying to push Linux to the low-end and by bashing Linux on high-end servers, portraying Linux as a naïve community of hackers rather than viable business people, and trying to position Linux against low-end Windows servers.
Words from the Net
”Linux is an irresistible force and, no matter what some dinosaurs over there may think, SUN is not an immovable object, it will orbit Linux or it will fall into Linux and burn up. Whether they survive the journey or not they will come. They are coming over even now.”
Linux Community Posting on the Internet
Linux will continue to advance on high-end servers and the success of Linux on zSeries mainframes proves that Linux is a viable choice in large enterprises. The success of Linux businesses such as Red Hat and SWsoft proves thus. We in the Linux community understand while Linux competes with Windows on the low-end, it also competes with Solaris in mid-to-large enterprises.