OpenBSD Journal

Web server comparison a b

Contributed by Dengue on from the jihad-du-jour dept.

Noryungi writes : "A company called has released a performance comparison between the most used free-UNIX. The BSDs seem to be on top of the performance heap, with the lion share going to FreeBSD. OpenBSD and NetBSD are almost always in second place, with Linux and Solaris (x86) at the end. Fairly interesting. OS Flame War at 11. "

(Comments are closed)

  1. By David Terrell () on

    This appears to be the presentation that was
    given at LinuxTag 2000 (Linux Day 2000) in Germany
    that has been discussed a lot since then. Glad
    to see the data show up somewhere. More specific
    data including the scripts or source code used
    in testing would have been nice -- ah well. Looks
    good. Interesting to compare how everybody's
    performance seems to go up with the FreeBSD NFS
    client instead of Linux's. :)

  2. By Tom Kyle () on

    the article seemed pretty well done, but I thought it strange that they hampered their test system with IDE and only 64mb of memory! Running a common, hi-performance SCSI controller would seem a bit more logical than a pretty new VIA IDE chipset.

    Also, I think it might be fairer to compare Solaris running on Sparc hardware. An Ultra 10/450 would be a fairly comparable machine, and let's be honest - Solaris is run on WAY more UltraSparcs than x86s. But it sure is cheaper to reinstall OS's on the same hardware. ;^)

    Finally, how 'bout mounting NFS volumes from a third-party fileserver, say a NetApp filer? I know, I know, cost issues...

    1. By Warp Eight Bot () warpeightbot at-sign yahoo! dot-com on

      P'raps he was trying to compare apples to apples, i.e. all OS's on same, commodity hardware (as in, you can just run out and buy one of these for cheap-ninety-cheap off the street, instead of going for the ubercool system we all know and lust after). Re-stated: "If you want to set up a server on the cheap, this is the kind of performance you can expect from the freeware OS's you're going to be choosing from." No, it doesn't address scalability, or getting the most from a system by going the hardware route. In that, I think it's useful; there are a lot of people out there who just want their pretty dog and pony show on the web, with an 800 number or an email to fax form at the bottom, no muss, no fuss. This is what is being addressed.

      You wanna hot rod it, vote with your feet and find a different study.

      1. By Tom Kyle () on

        "In that, I think it's useful; there are a lot of people out there who just want their pretty dog and pony show on the web, with an 800 number or an email to fax form at the bottom, no muss, no fuss."

        Er, so why would you be concerned about performance . If that's the focus group, ease of setup, security, and reliability are far more important than performance. Commercial support also becomes important to many of these people.

        If you're going to be out for performance, you're not going to buy a low-buck system. Unless, of course, you buy several and get them to distribute the load...

    2. By Nobody You'd Know () on

      64mb of memory is enough for those tests if the OS handles it even reasonably well. If it wastes memory or has poor buffer management, then it loses. That was probably the point; ANY system can get good performance with effectively infinite large main memory. Even Netware:)

      As for running Solaris on an Ultra, that's a nice idea, but it would change the study from "which OS is best" to "which platform is best." That wasn't the author's goal.

      And as for IDE vs SCSI, I suggest you get a modern IDE drive and controller, make sure there is nothing ATAPI on the same channel with the drive, and test it. Frankly, if your OS can do the fancy DMA modes and whatnot on the newer IDE stuff, you can blow away all but the MOST expensive SCSI controllers and disks at a ridiculous fraction of
      the cost. When the newer serial IDE spec makes it into real hardware, that'll pretty much be the end of SCSI on anything smaller than a hardware RAID array, save for a few religious zealots of the same sort who still swear that they can "hear" the sampling error on CDs.

      1. By Matthew Weigel () on

        As for running Solaris on an Ultra, that's a nice idea, but it would change the study from "which OS is best" to "which platform is best." That wasn't the author's goal.

        Yeah, but "which platform is best, particularly at a certain price point," is what matters -- if discussing what matters "wasn't the author's goal," then, well, why should we care what he says? It clearly doesn't matter. Particularly when you consider systems like Net and OpenBSD, where really can pick the platform to suit the needs, standardizing on a single architecture obfuscates the results. And as for IDE vs SCSI, I suggest you get a modern IDE drive and controller, make sure there is nothing ATAPI on the same channel with the drive, and test it.

        And I suggest you get two modern IDE drives on a single IDE controller, versus three modern SCSI drives on a single SCSI controller, and compare performance. I don't care how fast an individual drive is on my servers (beyond a certain point); I care about how well the disk handles increasing numbers of concurrent requests. Remember, we're not talking about workstation performance here, but server -- web and NFS, most notably -- performance. If I didn't mind not being able to switch out disks amongst my systems I'd be (mostly) happy to use IDE on workstations, since you're going to need less disk access. On the server? No.

        1. By Nobody You'd Know () on

          If you don't care about the intent of the study the guy did, then don't read it. Reading a study about primate hair coloration and then claiming that it "doesn't matter" and should have been about primate mating habits is a waste of time.

          Besides, bang for the buck, if you're spending less than $10,000, -quality- IA32 hardware wins. Any honest person who does the numbers and doesn't let technosnobbery cloud his sight knows that -without- a big new study.

          It sounds like you want to know what's good -above- that price range. Honestly, if you're quibbling over saving a couple thousand bucks on a machine that expensive, you either can't afford it or don't need it. This is why you don't see pricing information for E10k machines on Sun's web site. Nobody comparison shops for that stuff unless he's either too poor to buy it or else amazingly, painfully cheap. (Yes, there ARE corporate IT people who are that cheap. They pay for their foolishness daily, but they don't have to write up purchase orders for those expenses, so they seem happy. Let them eat cake...)

          By the way, IDE on motherboard controllers isn't quite the same beast as if you buy a decent one, and a decent one is still damned cheap. Yes, for large disk arrays, SCSI still has an edge. However, talking about large disk arrays on commodity hardware is an indication that you are either clueless or like wasting money.

      2. By Anonymous Coward () on

        Have you ever tried writing to multiple disks on the same IDE controller at the same time? Bad idea. BTW, you're out of your mind if you think people can't hear the difference between CD's and vinyl.

        1. By Nobody You'd Know () on

          Actually, what you're referring to is writing to multiple disks on the same cheapo integrated IDE controller, but seeing as your tone suggests that you're repeating something you heard somewhere but never bothered to really check out for yourself, I'm guessing you didn't even know there WERE any others.

          And yes, you -can- hear the difference with vinyl. Every scratch, every imperfection in the pressing, every extra bit of hiss added by the mechanical nature of the system. Some "audiophiles" are stupid enough to think that the last of those is "warmth" or some similar attribute that they claim CDs don't have. It is true that CDs don't sound like records - because they're an -accurate- reproduction instead of a shoddy approximation. However, I realize that there are people out there who've built hobbies, careers, and even egos out of denying this fact, and that they are not going to change their minds just because reality disagrees with them. In some cases, it is a sense of "the good old days," in some cases, a desire to maintain a feeling of elitism(the obnoxious part of "audiophiles," to be sure,) and in some cases, it is merely a desire to seem "interesting" or "different." Adherents of such religions are probably interesting to sociologists, but I'm sick of them, myself.

  3. By Dave Uhring () on

    It was quite obvious from the bonnie results that the Linux kernels had NOT been configured for DMA. Several months ago, I ran bonnie on OpenBSD 2.6, FreeBSD 3.4, Linux 2.2.14 & 2.3.46 and Solaris 7 & 8, and except for Solaris, the others used DMA for disk I/O and all were comparable to within only a few percentage points. There was no clear leader overall as that site's graphs indicate. Solaris on the Intel Architecture is, simply put, a pregnant sow where speed is concerned. This is quite understandable since Sun Microsystems wishes to sell hardware and allowing IA hardware to mimic SPARC hardware's speed is, of course, taboo.

    1. By Josh () on

      Linux DMA *WAS* used. If you notice in one of the last slides he mentions the use of hdparm just as NMBCLUSTERS for BSD was touched on. But softupdates were left unmentioned. *BSD handily beat Linux again.

      1. By Dave Uhring () on

        Well enough, then. Run your own comparisons. I believe _everything_ that someone else tells me...NOT!

        1. By Nobody You'd Know () on

          Linux -does- have some performance problems, as Linus will readily admit. Only in the recent
          development kernels are they starting to get a
          system that handles heavy load well, and there are still problems to be resolved. Yes, I think they'll resolve them. Even though I'm a BSD guy, I -hope- they resolve them. I think we all benefit when there are as many high quality systems available as possible.

          That's a point that seems to be lost on many people. Free software may be a competition of sorts, but it is not an economic competition. You don't "beat" the next guy by putting him out of business, and your continued survival is not at stake if someone else's code is better at one particular test than yours as of right now - so why do so many advocates sound like Baptist missionaries, out to save your soul? Really, really lame.

  4. By Alex Farber () on


    I am trying to install OpenBSD and get
    the error message "ne1: device timeout".
    Probably because my NE2000-card has io=0x300
    and irq=12, and not 10 how the boot message
    mistakenly says... Does anyone have an advice?
    Can I specify the irq at the "mdeia directives"-

    Thank you

    1. By Alex Farber () on

      Sorry - I mean of course, that my NE2000 card
      uses IRQ=5, since IRQ=12 is my PS/2 mouse.
      Everything works fine under Debian (potato) Linux.

      I have tried "boot -c" at the boot prompt and
      have been able to connect to the net. But after
      the installation I can not boot from the hard
      disk - everything stops after "ne1: .... irq 10".

      Booting from official CD's works fine however...
      Does anyone have an idea?

      1. By Alex Farber () on

        Maybe it's not my network card? I have
        tried booting with "boot -c" and "verbose":

        >>> probing for ne1
        >>> probing for ne1 succeded
        ne1 at isa0 port 0x300/32 irq 10
        ne1: NE2000 Ethernet
        ne1: address 00:40:34:26:6f:f2
        >>> probing for ne2
        >>> probing for ne2 failed
        >>> probing for we0
        >>> probing for we0 failed
        >>> probing for we1
        >>> probing for we1 failed
        >>> probing for ec0
        >>> probing for ec0 failed
        >>> probing for eg0

        Okay, irq 10 for my NE2000-card is wrong,
        but booting stops after ">>> probing for eg0".
        What could I do? Thank you


        1. By Alex Farber () on

          In tradition of replying to my own
          posts: I've solved that, thanks


  5. By Ragnar Beer () on

    I just recently switched from Linux to OpenBSD for security reasons. Inspired by this article I ran a couple of benchmarks today. My setup is: Apache 1.3.12 serving Zope 2.1.16 content via pcgi. Zope again for this benchmark doesn't serve content from it's own object database but from MySQL. The test machine is an Athlon 650 with 256MB. Accessing one record from the database Apache served 45.9 requests per second on Linux 2.2.14 and 49.6 rps on OpenBSD 2.7. However accessing 1000 records containing 10 Byte strings Apache served 25.2 rps on Linux but only 13.5 rps on OpenBSD. As a comparison: Linux on a K6-2/300 with 256MB: 16.7 rps. I'm glad I have to access such a lot of records at a time rather seldom only but I wonder how this breakdown in performance can be explained when OpenBSD is clearly superior to linux when accessing one record only. Any suggestions?

    1. By Jeff () on

      Is your database backend on your Linux server
      threaded, and on your OpenBSD box unthreaded?

      That would probably be one of the biggest
      losses (if its the case), otherwise mysql
      might have some optimization tricks on Linux.

      1. By Ragnar Beer^ () on

        I'm using the OpenBSD 2.7 MySQL package. I don't know how it was compiled but 'mysqladmin status' reports up to 6 threads under quite heavy use which doesn't appear to be much.

  6. By jdube () jdube on jdube the commercial version of UNIX did the worst. I wonder how BSD will fare when, in a few years, it is more mainstream versus new UNIXes, say, HURD or EROS (I think eros is unix, at least :)

    1. By Nobody You'd Know () on

      People said the same thing about a dozen other systems in the last 10 years, most of which you've probably never heard of.

      It is one thing to -talk- about your cool new system, or write a version suitable for a few guys to hack on here and there. It is quite another thing to produce something genuinely useful to a significant number of people. In short, no matter how promising a system seems, wait to proclaim it as a future contender for significant usage until it at least has some users who aren't also developers or testers.

      So far, "traditional" unixes have been around for roughly 30 years, and nobody has even come close to making a serious competitor in the eyes of anyone but developers of new systems and enthusiasts for anything and everything new.


Copyright © - Daniel Hartmeier. All rights reserved. Articles and comments are copyright their respective authors, submission implies license to publish on this web site. Contents of the archive prior to as well as images and HTML templates were copied from the fabulous original with Jose's and Jim's kind permission. This journal runs as CGI with httpd(8) on OpenBSD, the source code is BSD licensed. undeadly \Un*dead"ly\, a. Not subject to death; immortal. [Obs.]