Top Five Reliable Providers

X86BSD writes "Interesting survey at Netcraft showing the most reliable hosting providers for June. Interesting that not just the top 5 are FreeBSD but that the top 10 come from all variants in the industry."

Comments

I'd agree, but

[tyrani]

I'd agree that server reliability depends on the O/S used, reliability has much more to do with the installation and setup of the server.

So, congratulations should go out to the sys admins of those servers.

Re:I'd agree, but

[Motherfucking Shit]

rejected (16) accepted (0)
Is there a psycological term related to getting your stories rejected on slashdot?

"Nogirlphrenia," of course :)

Re:I'd agree, but

[Ziest]

Agree. Long uptimes are a recipe for disaster. 2 things can go wrong 1) the system on disk has changed under the system in memory. Broken or missing shared libraries and init scripts. 2) my fav, the disks stop spinning. This is lots of fun. Try it some time. Go to one of your older machine, the one with 500 plus days of uptime. Shut it down, remove the power cord, wait 3 minutes, reattach the power cord, put your finger on the power switch. OK, gentlemen, place your bets. Will all the disks spinup? If they do not, you get to find out how good your backups really are. When was the last time you backed that system up? When was the last time you verified the backup? A simple "Put more memory into this machine" becomes 3 or 4 days of living hell as you run around trying to figure out what happened, try to reconstruct the contents of that dead disk. Fun, fun, fun.

Anyone who allows a machine to go more than 30 days without a reboot is asking for trouble. There is a reason why mainframers have a maintenance window every Sunday.

Re:I'd agree, but

[gregmac]

Anyone who allows a machine to go more than 30 days without a reboot is asking for trouble.

I disagree. While 500 days is quite a few, there's no problem as long as you're diligent. Set up a script to do backups to another system on a regular basis (daily, weekly). I do incremental backups every workday at midnight using rsync. Which means if you mogrify a file, you have up to a week to get back an old copy. I used to have this go to a tape drive, and the tapes just had to be swapped out weekly. Now I just ocasionally archive to CD.

As far as disks dying, RAID1 helps, and is very cheap now and available on many mobo's. With an older server, maybe you don't have this, but maybe if it's a concern, its time to upgrade. Bigger servers often have RAID arrays already.

If you're making any major system changes - ie, the way things boot - then be sure to test it out. Be sure it boots after you make the change and are sitting right in front of it. I'll also assume you're not going to be making changes like this on a mission-critical production server, and actually testing them out on another system first. If you have a production server that you make major changes to without testing, I think it's obvious that you're asking for a disaster.

Re:I'd agree, but

[buss_error]

Agree. Long uptimes are a recipe for disaster. 2 things can go wrong 1) the system on disk has changed under the system in memory. Broken or missing shared libraries and init scripts.

You're the admin. You're supposed to check for this. If the system isn't all that important, I may add patches without checking them on a test system, but if it's important, no patches get added until they are checked on a test system.

2) my fav, the disks stop spinning. This is lots of fun. Try it some time.

You're the admin. You're supposed to be doing backups. Personally, if I think there's a good chance that the drives will fail when I'm doing something ( eg: greater than .5 percent) I make 2 back ups. Tapes can break. Also, I've not seen disks refuse to spin up with out powering off for a while (more than 5 minues). Frequently, you can get the disks spinning again by (gently!) tapping them with a screwdriver. If that doesn't work, sometimes heating them with a lightbulb will work. Heatlamps work too, but you need to be careful not to overheat the drive. I also try to get drives on critical systems replaced every 2 to 3 years. RAID helps here.

Keeping the network, hardware, OS, and applications up is important, but just as important is abuse response. There are a few hosting companies out there that do a wonderful job of keeping things ticking over, but fail absolutely at terminating abusive accounts. Hosting at one of these sites is inviting having your email blocked at the very least. Some sites block all traffic based on what's in the block lists. Part of due dilligence is checking the history of a host by checking at SPEWS, SPAMHAUS, SPAMCOP, News.Admin.Net-Abuse.email, News.Admin.Net-Abuse.Sightings, and other customer's experiences.

I can't find my link to the dead tree report I use to check out hosting companies at the moment, but there are several very nice writeups out there that focus on choosing a good hosting/co-lo company.

Re:I'd agree, but

[secolactico]

A simple "Put more memory into this machine" becomes 3 or 4 days of living hell

Ah, but the realist (pesimist) admin will always hope for the best and plan for the worst. Critical machines/services need to have standby hardware. If I need to do a hw upgrade/software patch on a production machine, I'll make sure the standby machine is up to date and working before touching the main production machine. That way, If I I can't bring the machine back up in 10 mins, I'll know I have a Plan B machine.

Same goes for any other network equipment. I've seen "carrier class" switches that simply decide to go south upon removal of a hot-swappable module (going by the book).

Re:I'd agree, but

[bcrowell]

I'd agree that server reliability depends on the O/S used, reliability has much more to do with the installation and setup of the server.
The end-user's experience is also much more likely to be influenced by how good their ISP is. Does their ISP drop the connection on long downloads? Does their ISP's modem bank get swamped in the early evening?

Re:I'd agree, but

[jdhutchins]

These are all problems that affect individual servers. If you have the money to keep a site up and running for 500+ days, chances are you have load balancing and more than one server. If one server dies, and you have it set up right, you won't have any downtime. You can also take time to do mantainance on one server, and the other servers can pick up the slack.

If one machine dieing brought down Google, they'd probably be constantly down with the number of machines they have. However, they have enough machines, and it's set up correctly, so that it doesn't matter if just one dies.

Re:I'd agree, but

[buss_error]

The point is, many home users push their machines. Sure, they are the admins, but they're also just regular schmoes.

Well, if the system is important to them, they should bestir themselves to learn how to make backups.

My biggest gripe about Novell and MS operating systems, and the Intell platform in general, is the inability to make a boot tape. BRU is able to make a boot disk that will allow a full restore from tape, but that function last I checked was only availble on (gag) SCO.

I want to throw a tape in the drive, tell the BIOS to load the OS from tape, and restore that sucker to disk. I don't want to have to do partial installs, booting from floppy/CD, or any of that crap. Load and go is what I want. Once you've been able to restore a full system image from tape, you'll wonder why all vendors don't offer that function.

Before the *BSD is Dying trolls start...

[Anonymous Coward]

According to Netcraft,

Intriguingly, all of the Top 5 placed sites run the FreeBSD operating system

I'm curious over the choice of the word "Intriguingly." My experience with FreeBSD has shown it to be nothing but rock-solid as a server OS. I actually prefer it over Linux these days (I was a RH-zealot for a couple of years until I "saw the light," as it were).

What would be intriguing were if Windows had nabbed the top 5 spots...

Re:Before the *BSD is Dying trolls start...

[Alan Cox]

I don't see anything intriguing there. The Linux clock wraps at 497 days. It's also not "intriguing" as such because FreeBSD is an extremely stable OS.

I am suprised AIX didn't show up in the top five I must admit

Re:Before the *BSD is Dying trolls start...

[SoSueMe]

In the bottom 10 (of 50), 1=HP-UX, 2=Linux, 3=Windows, 4=Solaris.
I think it is more the admins, rather than the OS's.

Re:Before the *BSD is Dying trolls start...

[qortra]

I guess I'll have to disagree with both you and AC. I think it's intriguing if any particular OS nabs all top 5 spots; I (and apparently the folks at Netcraft) imagine there would be more variety at the top. There are many other very stable OSes out there, such as some flavours of GNU/Linux (read "Debian-Stable"). But all that to say that I'd rather not nitpick Netcraft about one particular word. In the past they have chosen to put a humorous twist on server uptime data (an otherwise dry topic), and I have always liked the result.

Re:Before the *BSD is Dying trolls start...

[swb]

I am suprised AIX didn't show up in the top five I must admit

Are there hosting providers using AIX in their hosting environments? I would think that RS6000s would be just too expensive in comparison to blades or generic 1 or 2U x86s for hosting environments.

I'm sure there's some popularity in ASP environments where you're providing an entire application (interface, DB, logic, etc), but for basic hosting it sounds like it'd be unaffordable to use RS6000.

Re:Before the *BSD is Dying trolls start...

[eht]

Blame everything on the Linux clock wrapping at 497 days, well you might want to have that fixed eh?

I'd like it fixed so it can stopped being used as an excuse.

Or you could read the article and find it has nothing to do with anyone's uptime clock, it's by failed req% in the month of June, but that would be too hard.

Re:Before the *BSD is Dying trolls start...

[Tokerat]

The Linux clock wraps at 497 days.

For those of us who are slightly less familiar with Linux (heh), could you explain (or provide a link to a page explaining) why this happens, and what the reason for it still happening that way is?

Honestly not a troll or anything, just curious and I figure I can get an accurate answer from someone like Alan Cox...

Re:Before the *BSD is Dying trolls start...

[Admiral Burrito]

*BSD is dying

As usual, slashdotters aren't reading the article.

The article says that BSD-based services are very reliable, and tend to not die. And it is a Netcraft article, so...

Netcraft confirms: *BSD is not dying.

Re:Before the *BSD is Dying trolls start...

[FattMattP]

But he's not the one using this "bug" as an excuse.

An excuse for what? Netcraft checks your uptime by pinging your machine. They aren't going to be able to log into your machine and check the internal counter that keeps track of your uptime. The limit in the Linux counter has no relevance here.

Re:Before the *BSD is Dying trolls start...

[nacturation]

An excuse for what? Netcraft checks your uptime by pinging your machine. They aren't going to be able to log into your machine and check the internal counter that keeps track of your uptime. The limit in the Linux counter has no relevance here.

Wrong! From the Netcraft page:

Which operating systems provide uptime information ?

Operating systems we can usually work out uptimes for are:

• BSD/OS
• FreeBSD [but not the default configuration in versions 3 to 4.3]
• HP-UX [recent versions]
• IRIX
• Linux 2.1 kernel and later, except on Alpha processor based systems
• Solaris 2.6 and later
• Windows 2000
• Windows Server 2003
• Windows XP

Operating systems that do not provide uptime information include;

• AIX
• AS/400
• Compaq Tru64
• DG/UX
• MacOS
• MacOSX
• NT3/Windows 95
• NT4/Windows 98
• NetBSD/OpenBSD
• NetWare
• OS/2
• OS/390
• SCO UNIX
• SunOS 4
• VM

Additionally HP-UX, Linux, Solaris and recent releases of FreeBSD cycle back to zero after 497 days, exactly as if the machine had been rebooted at that precise point. Thus it is not possible to see a HP-UX, Linux or Solaris system with an uptime measurement above 497 days.

Why do some Operating Systems never show uptimes above 497 days ?

The method that Netcraft uses to determine the uptime of a server is bounded by an upper limit of 497 days for some Operating Systems (see above). It is therefore not possible to see uptimes for these systems that go beyond this upper limit. Although we could in theory attempt to compute the true uptime for OS's with this upper limit by monitoring for restarts at the expected time, we prefer not to do this as it can be inaccurate and error prone.

Hmmm...

[Dark Lord Seth]

I don't know, I think http://localhost/ is the fastest, most reliable of them all instead. Shit content though...

What about...

[Mensa Babe]

...SCO?

Re:What about...

[archen]

Actually I'm interested in SCO too. Having had the misfortune being an admin for a SCO system for 2 months (before we switched to Linux) I wonder if anyone seriously would use SCO as a webserver. If s o I'd really like to hear about their experiences =P

Netcraft has all sorts of interesting data to dig through. But no place just too look for a total list of OS hosts. Also sort of neat to see what domains interest people the most. (Like #5 linuxsucks.org running on Linux)

Fasr? Reliable?

[phud]

Looks like netcraft needs to switch

Re:Fasr? Reliable?

[Neophytus]

Netcraft says Netcraft is using FreeBSD

Re:Fasr? Reliable?

[kilonad]

Netcraft says Netcraft is using FreeBSD

No wonder Netcraft is dying... ;)

While we're on Netcraft...

[Anonymous Coward]

This is a pretty interesting link. Apache is great and all, but the dark side works with FrontPage.

BSD

[grub]

I'm a happy {Free|Open}BSD user after switching from Linux about 4 years ago. I haven't had to monkey with Alice's patches to Bob's kernel mods to run Charles' software since. This isn't a Linux slam but with the BSD family, once you have a stable system it just runs until the hardware dies. In fact I only reboot my OpenBSD boxes when there is a security hole (you know how often that is!) or big upgrade to the kernel/OS that I want, not just the Kernel du Jour.

Re:BSD

[Moridineas]

I'm in the same situation. Use Open for Firewall computeres, Free for everything else. Stable and easy to maintain. The easy to maintain part is about the most important to me.

Re:BSD

[IamTheRealMike]

Isn't that true of *any* solid server OS? If you get it set up correctly then leave it, most machines will run until the hardware dies.

I personally don't know many real Linux production servers (as opposed to bobs personal box) where the admins mess with kernel patches - ever. They normally use a stable distro, normally Debian or one of the older Red Hats, and just leave it.

Re:BSD

[TheRaven64]

In fact I only reboot my OpenBSD boxes when there is a security hole (you know how often that is!)

Yes, we do. They can't even go 7 years without a root hole in the default install. Pathetic ;)

The thing that made me switch to *BSD, however, was not the security or the stability. It was the documentation. Nowhere else have I found such a wealth of (well written) documentation as I discovered on installing FreeBSD.

Already slashdotted...

[Trurl's Machine]

And I am supposed to take *they* word on what is the most reliable technology?

OS is a commodity

[nuggz]

Although FreeBSD made the front page, it looks like the others are also represented.

To me this suggests that they are all capable, and the differences come from somewhere else, the setup and administration.

Bzzt

[Motherfucking Shit]

If you look at the chart, 2 of the top 5 ARE running Windows 2000.

I presume you're talking about this chart (the one linked in the story doesn't show OS), which lists the top hosting providers over the last 24 hours... Not for the month of June.

In any case, I'm a bit skeptical of the data. They seem to be monitoring the providers' own websites, not their clients' machines or sites. For example, the 24 hour chart shows Interland listed as Win2K... That may be true of www.interland.com, but most of the Interland clients I know are either running dedicated *nix boxen, or running off Solaris virtual hosting accounts at Interland's Communitech branch.

Regardless, I certainly wouldn't rank a host based on their ability to keep their site up. Most if not all of them serve their corporate site from a server unrelated to their clients, and the site (and server itself) are rarely messed with. This is especially true with shared/dedicated hosts.

Machine footprint

[Gandalfar]

With footpring changing tactics I'll bet well soon see increase of Dreamcast based providers.

Newsflash in not so far future: "IIS down 5% Dreamcast up 15%"

Wanna see something REALLY funny on Netcraft...

[Lobo]

http://uptime.netcraft.com/up/graph?site=www.sco.c om

top 50 are typical

[jd142]

If you look at the top 50, you get these percentages:

20% FreeBSD
26% Windows (NT and 2000)
30% Linux
22% Solaris
2% HP-UX

This is fairly close to the overall distribution of servers. It usually works out to about the same numbers. Currently, Apache is at 63 percent and IIS is at 26%. Which would be about right if all of the Windows boxes are running IIS and most of the unix variants are running Apache.

So the news appears to be that the top 50 most reliable providers are, generally speaking, reflective of the whole of all providers. Which means that it isn't just the server os that makes a hosting company reliable, it's the hardware and the techs. There's no magic bullet for uptime. You can't categorically say that one os is the absolute best. You have to include the technical skills of the admins in the equation.[1]

[1] You *might* be able to state that free/open source software is more easily secured, but I suspect that the admins running those 25 bsd/linux sites would tell you that their skills made a difference in their uptime.

This is not surprising.

[Krapangor]

Main difference between FreeBSD and other system is not it's very good stability but the professionalism of the user base.
FreeBSD comes from an academic background and has much more high-profile users than any other system.
Even the very stable Linux system is dominated by hobbyists. The default installations of non-*BSD system are usually feature laden and sometimes broken. And note that stability of the kernel is not the only issue. If you fuck your configuration then you are fucked for good. It's a common misconception that a stable kernel leads to a stable system.
So, the pros and PhDs tend to use FreeBSD, not only for the above mentioned issues but also due to the clean design, tight codebase and modern algorithms. Note that e.g. FreeBSD was the first system with O(log(n)^2) swapping. This gives a double advantage: you get a stable system with a high-profile userbase. That's why we will always see FreeBSD on the top.

Re:This is not surprising.

[swb]

My feeling is that the main difference between FreeBSD and Linux distros is that FreeBSD is a complete *system*. Linux too often feels like a GUI glued onto an operating environment that itself is a kernel glued onto a bunch of utilities.

FreeBSD seems like somebody paid more attention to the components, and a good number are unique to FreeBSD and not GNU parts. Even the contrib aspects of FreeBSD (gcc, sendmail, etc) are well integrated and not just bolted on.

Re:My trials with *BSD

[fmaxwell]

It sounds like you don't know how to properly configure the OS.

I agree completely. Whether someone likes BSD or not, its just assinine to assume that 20-minute-plus times to copy a 17MB file are normal on a PIII system. It's pretty damned obvious that BSD does not have slow file I/O when you consider that BSD variants are the OSs of choice for major hosting providers, massive commercial databases, Yahoo!, etc.

No particular order

[nuggz]

And if you look at the data you see.
The top 40+ have no failed requests, and it is just minor differences in response times, and it isn't overly clear if they are even sorted by that.

rant mad libs....

[RiscIt]

Let me save everyone some time...

"[name of host I use/used to use] should not have been in the top [number 50 or less]! I used them for [length of time] and I had [random number] problems over a [1-12] month time span! They are TOTALY incompetent and I wouldn't wish them upon my neighbor's dog's fire hydrant!"

Liars!!

[borgdows]

Hey, some sites of the Top 10 are lying about their OS!

# www.nyi.net - FreeBSD
ok

# www.about.com - FreeBSD
ok

# www.nac.net - Windows 2000
ok

# www.interland.net - Windows 2000
LIAR!!
nmap -O www.interland.net gives :
TCP/IP fingerprint:
SInfo(V=3.00%P=i686-pc-linux-gnu%D=7 /13%Time=3F116 9B0%O=80%C=-1)

# www.inetu.net - FreeBSD
ok

# www.jumpline.com - Linux
ok

# www.myhosting.com - Windows 2000
LIAR!!
nmap -O www.myhosting.com gives :
Remote operating system guess: AIX 4.3.2.0-4.3.3.0 on an IBM RS/*

# www.expresstech.com - Windows 2000
ok

# www.hostopia.com - Linux
ok

# www.verio.com - Solaris
ok

There's something rotten in Netcraft kingdom! ;)

Re:Liars!!

[Ascender]

You say that, but I believe that netcraft detects the OS based on responses to queries sent to the webserver.

Nmap, on the other hand, detects the OS based on the random-number sequence generation of TCP packets.

How does this affect things? Well, if you have a load balancer or firewall that forwards the HTTP connection to another webserver somewhere, then nmap will return the OS of the firewall or load balancer, whereas netcraft will return the OS of the final webserver.

Re:Liars!!

[borgdows]

Possible, but the HTTP OS header is a lot more easier to change (especially on Apache) than the seqence of TCP packets.

its been /.'ed

[Adler]

obviously they don't use one of their most reliable hosting companies...

Re:Win2k

[Mostly a lurker]

if a hosting company is using Win98 as a production hosting server they should be prevented from breeding.

Do not worry. If using Win98 as a production hosting server, they would be too busy problem solving to have time to breed!

You missed the point

[taxman457f]

If you read the article you would see:
A summary showing the ten providers whose sites experienced the fewest failed requests and the fasest connection times during June

So this one is about performance, not uptime.

So the fact that Freebsd tops out on performance *and* uptimes is pretty amazing. What I will give you that you are correct about, is the performance is highly dependent on the hardware and the skill of the techs. Uptime is not everything for sure.

"There's no magic bullet for uptime." Well if it *is* uptime you're looking for, the closest to a magic bullet you can get is FreeBSD or BSD/OS. 5 years is a damn long time.
essentially the top 50 uptimes ever, is fully dominated by BSD
http://uptime.netcraft.com/up/today/top.last.html

AIX for web hosting

[macwhiz]

Are there hosting providers using AIX in their hosting environments? I would think that RS6000s would be just too expensive in comparison to blades or generic 1 or 2U x86s for hosting environments.

I worked at a major telco ISP that used AIX. In fact, not only were the web services hosted on AIX, but they were hosted on a RS/6000 SP2 parallel supercomputer.

This sounds like overkill, and in some ways, it is... but the SP2 is, in essence, a very fancy rack. Each SP2 frame has a number of nodes in it; each node is a self-contained RS/6000 system. The major difference in an SP2 node is that it has a SP Switch connection. The SP Switch is a very high speed switch fabric that allows the nodes to communicate. Combined with heavy-duty software and fault-tolerant design, you wind up with a parallel supercomputer...

...or one heck of a rack system that lets you run the nodes as individual servers, or in clusters, with lots of bandwidth and control.

For a service provider that wants to lure people in with a low starting cost, and hope that turnover from downtime isn't too bad, AIX can be expensive. I used to dislike AIX, because of its reputation as "not quite UNIX." Once I had the opportunity to use it, I found that it really is well suited to many ISP tasks. AIX has inherited a lot of attitude from IBM's mainframe days. IBM's mainframes were used in "can't go down" environments. AIX has many features that share that design philosophy.

As for the cost... as with any major manufacturer system, there's the published cost, and there's the cost you negotiate. If you are buying a whole setup, you can usually cut a deal. Of course, if you're buying major manufacturer equipment, you're already committed to paying more than you would for a white-box open source system, presumably because you want advanced features that haven't made it into OSS yet, or you want support.

(I've found that IBM's AIX support kicks ass. When I'd call Sun, even with a Platinum contract, I'd usually get someone who'd do the same SunSolve search I already tried, then promise to get back to me some day. Calling IBM gets results... they will put as many people into conference as they have to in order to get enough subject-area experts talking to figure out the problem and resolve it, preferably on the same call. A far cry from "RTFM and then post to the mailing list!")

The only ISP task that I found AIX had trouble supporting was INN. At least at the time I was working with it, AIX had resource limitations that caused trouble for very large INN installations. (This ISP was working with a two terabyte news spool.)

Looking at the whole 50

[LooseChanj]

I see the lowest freebsd is #41. The lowest Solaris 8 is #49. Lowest linux is #48. Lowest Win2k is #47.

Those are the 3 OSen which comprise the top 10, which has 5 fbsd, 2 Solaris 8, 2 Win2k, and 1 linux.

Totals for the top 50 are 10 Freebsd, 15 linux, 8 Solaris 8s, 2 'Solaris', 12 Win2k, 1 NT4/Win98 box admin'ed by a crack smoking monkey, and a lone HP-UX. (I know missed one somewhere, but screw it...I'm not recounting.)

Now, what does this tell us? There are FreeBSD users at the top and bottom. Same for Solaris 8 and Win2k. Linux too. OS doesn't really seem to be much of a factor. Hardware and network reliability I would expect to be more relevant.

My conclusion? The people who chose these things, along with the OS, and setup and maintain them, to land themselves in the top 5 must have made better decisions than the rest. That the people who chose the most reliable hardware and networks also chose FreeBSD...well, it only goes to show. :-)

Re:Looking at the whole 50

[Mastoid]

No matter what you say, *someone* will disagree.

That's not true.

Operating system has nothing to do w/ reliability

[chrysalis]

IMHO the reliability of an ISP has nothing to do with the backend operating systems, especially when the study only considers the OS of the public web servers.

I'm sure that even a Win95 based ISP can provide a very reliable service. It's only a matter of redundancy.

On the other hand, the company I'm working for runs FreeBSD on its web and mail servers, but thanks to the dumb way things are installed and the lack of redundancy, a global uptime of 24 hours would be an all-time record.

With no possible single point of failure, with load balancers and correct usage of protocols like HSRP, service can be guaranteed even if some servers are continously crashing.

Have you ever seen Google unreachable? I've always seen it up. Although Google runs Linux. But they have properly designed their network for high availability. In an old Slashdot article, there was an interview of a Google techie who explained that if 1, 2 or 100 servers were down, it would have absolutely no impact on the service.

So at least for ISPs, I really think what matters is the skills of the network administrators. It brings another question : does the skills actually depends on the operating system they use?

Maybe. At least when you read mailing-lists of different operating systems, you can clearly see some common interests of the related subscribers. _This_ is really what makes differences between free operating systems. When it comes to reliability for traditional ISP services, either OpenBSD, Linux, FreeBSD or even Win2000 are quite comparable nowadays.

Re:Hmmm... Localhost... which one?

[frostman]

Hilarious!

I clicked on the localhost link just for the heck of it, since I had forgotten which project was running on my laptop's Apache server...

And MozillaFirebird helpfully took me to:

http://www.localhost.net.au/ ...which is, in fact, a hosting provider!

Ghosts in the machine, I swear.

Too Close to Call

[GoRK]

When you start getting into the territory of best uptime and fastest response between all network providers in the entire world, the game simply becomes too close to call, and the ordering of the hosts becomes simply the luck of the draw (that's why not one month of this set of NetCraft statistics is very consistent with any other month) Without monitoring statistics from both inside and outside for all these networks, it is virtually impossible to rank them all.

Just because some provider responds to a DNS or web query 1ms faster than the next guy doesn't mean they are a more reliable provider. Maybe they just have fewer customers and, thus, a little bit smaller DNS database or more room to 'spread out' the users between different machines. I know netcraft monitors from various geographical and network-isolated locations on the Internet, but the law of averages doesn't help them much here. Say they monitor from 100 hosts and one of those 100 hosts has a "connection reset by peer" type TCP (RST) error due to a local router (ie not the provider's problem) doing a DNS query and it logs a time of 100ms instead of 15ms and one failed attempt? It might not happen to the other guy's host monitored even 200ms later.

Netcraft would have to be monitoring from very large numbers of machines (100,000+) to even come close to being able to tell the difference between these networks from outside. I have a feeling most all of them on the list are very good and reliable, and aside from uptime, reliability means more than the response time of your DNS or web servers. No matter what OS or network gear you use, when you introduce redundancy and failover, you necessarily introduce more equipment and more complexity that will slow things down an inconsequential amount at the gain of another '9' or whatnot.

Anyway, the point is, don't take the order of this list too seriously. The fact that any company is on the list means they have made a serious effort to provide a good and reliable network.

~GoRK

Re:Too Close to Call

[josepha48]

If you look at this data over a period of time, like a year, you start to see a trend. I'm not sure what that trend would be, but, when you start to think about whose running what, (yahoo = FreeBSD) and then you add in cost (FreeBSD = free), you can then conclude that FreeBSD is just as good as Win2k. Thus if you wanted to cut costs in your company, using FreeBSD would be a good way to do so. Solaris and Linux also do pretty well. Using Solaris "could" be a more expensive solution than Windows if you buy a bog Sun box (yes I realize you can run Solaris on intel and get a small server for $2000).

The second issue that comes into play is what are they running on those OS'es? IIS on Win2k or Apache? Probably Apache on most UNIXish OS'es.

Then next it also must be determined things like what programming languages are they using php, jsp, asp.net, cgi, mod_perl, what? This also affects performance, as some languages are more effecient in the long run.

Lastly the number of servers installed and size. 1 or 500 will affect response time?

Yes I agree, this does not supply enough information, but it does seem to indicate more that ANY OS will do the job. The decision should be what language do you want to program in and what hardware do you want to use.

Hardware, anyone ?

[fmedio]

Interesting figures. But they don't say anything about the kind of hardware behind the OSes and different http servers. Nor do they describe the network topologies, routing policies or load balancing strategies used by the happy admins of the top-10 uptimers.

Yet, there is that embarrassing all-BSD top five. Tho I don't know how BSD or any OS can be of any help when you lose your storage subsystem.

Re:amen to that

[Bruce Perens]

Change the Cat5?!?!!? Why aren't you running heartbeat? You need never drop a single packet.

Re:amen to that

[DrSkwid]

well I didn't mention that the servers are also physically distant so the mirror is actually where I am, 10 mins from the co-lo.