The Practical Value of Mainframe Linux
|There have been a flurry of articles published recently about IBM's System/390 version of Linux. After reading those papers, I was left wondering why IBM has gone to such trouble. Here, we take a look at where Linux on the mainframe would be most -- and least -- useful. |
In S/390: The Linux Dream Machine, Scott Courtney takes the mainframe-as-dinosaur analogy in the opposite direction. He invokes a herd of forty-thousand raptors to describe the celebrated event of a single S/390 mainframe running over 40,000 simultaneous copies of Linux.
While that demonstration did highlight the robustness of the S/390's ability to subdivide itself into multiple virtual machines, it was also something of a parlor trick.
In practical terms, one 40,000th of a fully-configured S/390 will give each virtual "Linux box" about one 3400th of a CPU. These are fast processors , among the fastest available, but with a 0.2 MHz fraction of a CPU per Linux environment, this very efficient processor will not compare favorably to a whole chip of even ancient vintage.
The value of this capability, of course, is more in the ability to run several different operating systems, multiple versions of the same OS, and for isolating critical environments from each other. Using the S/390's virtual machine feature to create large numbers of small web server environments, for example, would not be economically advantageous. It would require many 12-CPU mainframes to support the same load as, say, a thousand 256 MB, 600 MHz Pentium III systems. Workloads that divide easily and don't individually require huge resources do not showcase the strengths of the mainframe.
While the thousand PC's described above are almost certainly overkill for 40,000 typical websites, I chose that high number for a reason; the cost of a fully-configured S/390 is roughly that of a thousand inexpensive Intel systems. That is the price/performance standard that the mainframe would have to meet to be cost-effective in this large, but simple, example scenario.
(Note that this is not a practical configuration; for the workload described above, one or two hundred PC servers would be adequate and would cost a fraction of the price of a single mainframe. To handle the load comparably, of course, several mainframes would be necessary.)
"Ahh, but what about large tasks that do not subdivide so easily?" is the obvious next question. Here, the cheap PC solution does not fare so well. If we switch from the single-CPU servers to four or eight-way Xeon servers , we make up a lot of the performance difference at the cost of some of the PC's price/performance advantage. Even so, there are some workloads that a mainframe can handle but an Intel eight-way SMP server simply cannot.
The high-end PC server, however, is not the main competition for the mainframe here. In this space, the large Unix system is the S/390's natural challenger. Sun's Cray-designed Enterprise 10000 , upgradable to 64 processors and 64 GB of memory, has dominated many of the large system benchmarks for the past few years. IBM's own RS/6000 Model S80 has recently displaced the Sun 10000 on several benchmarks. Unisys is also threatening to become competitive at the high end with their 32-way Intel-based ES-7000 but I know of no Linux plans for that machine. While two of these three machines have already shown a large performance advantage over S/390's on database-server benchmarks, the mainframe may outperform them all when running Linux. Linux is optimized for single-cpu to low CPU-count multi-processor systems. As such, it will likely run best on a relatively small number of very fast CPU's. While Sun's second tier system, the Enterprise 6500, "came in only a few percentage points behind" the top-scoring S/390 on the Peoplesoft Payroll Application, it required 28 processors to do so. [Although the Sun 10000 supports over twice the CPU's of the 6500, it is considerably less efficient, and would probably not perform as well as a 6500 if running Linux.]
Assuming a quality port of Linux, the S/390 will likely win the high-end performance contest.
In some discussions the issue of the S/390's "five 9's" reliability is brought up. However, IBM's 99.999% uptime claim is for clusters of mainframes, not a single system. A Parallel Sysplex configuration of S/390's can be very reliable indeed. However, I have seen no such claim for the reliability of a single mainframe. Extremely reliable systems can also be built using several other types of computers with clustering technology. In any event, individual large Unix systems are also extremely reliable -- when redundantly configured -- but the mainframe may still hold an advantage here.
At first glance, a single mainframe would appear to have a substantial manageability advantage over multiple PC servers. However, the advantage is certainly not proportional to the ratio of PC's to mainframes. Ten to twenty PC servers, identically configured with Linux, for example, can be quite manageable. Realistically, however, the smaller size of the PC's and an unruly workload may dictate that some of the PC's be set up differently. That leaves only the large Linux-capable Unix systems to compete with the S/390 in the realm of manageability. Therefore, while large centralized systems surely have a manageability advantage over higher numbers of smaller systems, it is not a characteristic that the mainframe can claim exclusively; the S/390 and the large Unix systems share the win on this point.
When top-notch, single-system performance and reliability are worth the mainframe price penalty, and when Linux is required -- and another Unix just won't do -- the S/390 might fit the niche. This may prove to be a narrow market, however.
More realistically, if your shop has an S/390 with some excess capacity, you can use it now rather than letting it sit idle. Then, as your OS/390 workload grows, you can migrate the Linux workload to another system. Due to the large increments in which mainframe capacity is purchased, and because systems must be configured for the rare, peak loads, most mainframe shops have a lot of free processor cycles. That unused capacity, along with the fact that nearly any Internet service can run happily under Linux, produces a scenario where the Blue Suit and the Tux go well together.
Other articles by Ray Yeargin
See all of Ray's abstract art prints on canvas.
|mail this link | score:9616 | -Ray Yeargin, April 2, 2000 (Updated: November 22, 2005)|
Articles and website reviews are owned by their authors.
The rest is © 2000-2012, Ray Yeargin.
Abstract Art Gallery