Slashdot Mirror
NT vs. Linux: Again
Jeff Molloy writes "The results are here
link " It's a
shame Linux didn't win, but it looks like the tests show where Linux might
have some deficiencies. Overall, it looks better than the original test, though.
← Back to Stories (view on slashdot.org)
If you can't help program then go out and test all this new stuff and send in bug reports. Let's have Linux set the standard again. It seems like, acording to the article, it was this way once and we lost it because Microsoft has pushed the bar a little higher and we lagged behind.
It's so nice of Microsoft to pay for this apache advertising. Just as a point of reference, 1800 hits/sec is the same as 155,520,000 hits/day. I think that it's safe to say that noone in the world gets more than 150 million hits per day of static content. Wait, there's a better way:
1800 hits/sec * average 2k/hit * 8192 kbits/kbyte = 29,491,200 bits/sec, or 29.5 MBits/sec. What's that now, a T3 line? I know that a T1 line is 1.5 MBits/sec. Ok, so apache on one of these boxes can fille the equivalent of 19.6 T1 lines by itself. If (a bit more realistically, how many 2k files get those types of hits) those are just 10k files (let's not get into pictures), that's 147.5 MBits/sec, more than filling a T3 line, IIRC, and definitely filling up aapr. 98.3 T1 lines.
What's the problem with Linux/Apache, now?
May I suggest, if you can afford this sort of bandwidth, that you buy one of those 32CPU sun E10000 servers and call it a day? (or a server farm of linux boxes, since you're serving up static files.)
Oh, if you're serving up >1800 files per second of 2k files, who are you?
Oh, one more thing. If this is all on an intranet, you'll still need Gigabit ethernet if you're serving up the 10k+ files, so the sun box still applies to you.
They laughed at Einstein. They laughed at the Wright Brothers. But they also laughed at Bozo the Clown. -- C. Sagan
From CMP's "Information Week" June 21, 1999:
"When every minute of downtime can mean millions of dollars in lost revenue, companies generally rely on applications that run on OS/390, Tandem NonStop Kernel, Digital OpenVMS, or Unix operating systems. But Windows NT is increasingly being deployed... so IT managers must find ways to increase the availability of their NT environments. To do it, they're adopting products and services that promise to provide extra protection..."
" 'Any system with lag time is unacceptable for running the application' says William Harris, NT Administrator for the Ohio Utilities. 'Money wasn't even a big deal. I's rather get quality and reliability and availability'. The organization...paid $75,000 to implement the (third party protection) system.
Translation (for those who need it): Management is telling IT they have to transition to NT. IT says, in order to be stable, we have to add third party help. Management says: "Here's a blank check."
It goes on to say that Unix, w/o third party software or service achieves "availability in the 99.9% range, as opposed to 97% for NT."
Now, what's the difference to a business between 97% and 99.9%?
IBM's NetFinity Availability Program guarantees 99.9 w/ NT. Cost: $220,000.
HP Mission Critical guarantees 99.9 with NT for a mere $300,000.
Imagine going to your boss and saying "Hey, how'd you like to save $300,000?"
JL Culp
Business Technology Consultant
Chair, LPSC
NT has put many services in kernelspace and has largely bypassed their HAL in favor of multimedia performance - especially video.
NT uses a multithreaded process model for IIS and SMB file-services that results in higher throughput but less stability. A single thread of the main process may die without completely destabilizing the server but if the main process dies then all child threads die.
Linux divorces the graphical user interface from the kernel thus ensuring stability (framebuffers are available for video enhancement though) and implements most services as userspace daemons.
Linux uses the forked process model to provide services to multiple users. This modem achieves stability in that if one process dies, the others continue as if nothing had happened. Both Apache and SAMBA operate in this way I believe.
NT has chosen performance over stability.
I believe that with kernel enhancements and profiling, any bottlenecks in the networking system can be eradicated causing Linux to perform much faster and possibly even beat NT in tests such as these.
Codifex Maximus ~ In search of... a shorter sig.
But here is our chance to show the world [and MS] why Linux and other OSS projects are such a good idea. By quickly implimenting fixes to the problems brought to light by these test, we can prove how much better OSS is.
Proposal: Annual or semi-annual benchmarking of NT [or the current MS server platform] and Linux [and any other OS's that want to compete I suppose]. By doing similar tests regularly, we can show how efficient OSS can be at fixing current shortcomings [as if 24hr bugfixes aren't enough].
Just a thought. ;)
BTW: Sorry for the overuse of the "OSS" buzzword
I'd help with implimenting fixes myself, but I'm not exactly an expert coder [I don't think "Hello World" will help Linux beat NT]
Ender
If at first you DO succeed, try not to look astonished!
Nothing to see here
But I don't think that this means that speed for web serving should be any more important. Getting back at Microsoft is not a reason to improve Linux in my book. There is are many other fronts that Linux heading toward like the desktop, embedded devices, and hand helds. I can imagine that if Linux is tweeked for web serving more than normal that some test will find Linux useless for embedded devices or something else that is important.
Microsoft right now sees Linux as direct competition as a server. It will be nice to see Linux compete back but don't expect NT to stand still. There are other servers also. How does Linux compare to Mac OS X?
And no more excuses. Linux is not the fastest. Deal with it.
For now.
--
I find these studies inadequate as data to inform a purchasing decision. While MS will claim that they have proven NT to be better than Linux for Web and file serving in the general case, I disagree. Here's why:
These studies do not address price/performance. P/P is one of the most important metrics in making a purchase decision; these studies measured only peak performance. That the prices of the Linux-based and NT configurations tested are not given indicates to me that Microsoft wishes price to be disregarded as a factor in purchasing decisions. To do so would be an irresponsible act for any purchaser. Consider that NT license fees increase dramatically with number of clients, while Linux's price is constant and lower than any NT option.
These studies do not address options such as clustering. Clustering is a common solution to the problem of constant high client load. It may well be a better solution (in P/P and in peak performance terms) than simply boosting processing power with multiple processors. It also has reliability advantages.
These studies are not generalizable to other hardware configurations. While MS will claim that they prove that "NT is faster than Linux" inherently, they do not. The HW configuration was selected for the first Mindcraft study, which has been proven to have been engineered to favor Microsoft. Hence the hardware configuration itself is suspect. An across-the-board comparison on various configurations, with P/P as well as peak performance measured, would be a more reasonable comparison of the virtues of the OSes themselves, and would also highlight particular combinations of HW and SW that are worthy of consideration for purchase.
These studies do not address security. The release version of MS IIS has outstanding security holes, including the recent one disclosed by eEye. This was a root compromise which took eight days for Microsoft to admit, and two more to fix. Microsoft classically avoids the subject of real-world security, preferring the proven-worthless tactic of security by obscurity. Security, of course, is a major consideration to be made in purchasing.
These studies do not address stability. Stability, like P/P, is an important metric for purchase decisions. It helps one determine how expensive a system will be to maintain -- one that requires regular resetting or reconfiguration in order to keep operating will cost in manpower; one which crashes a lot will cost in downtime. Downtime costs money in an enterprise situation, and hence should inform purchase decisions strongly.
These studies do not address changing real-world needs. A real server system is rarely left serving static Web pages forever. When needs change, performance will likely change as well. Building a system to meet a single, narrow-minded need is likely to lead to a dead end in terms of scalability.
These studies demonstrate nothing about the future. Based on past trends, one can expect the situation for Linux-based OSes to get better and better. The next version of Windows NT will likely offer decreased performance on the same hardware (due to increased resource consumption by the OS itself) whereas future versions of Linux will likely improve performance. Buying heavily into Windows NT leads one to platform lock-in which may damage one's ability to escape the expensive effects of bloat.
In short, I do not believe that MS has demonstrated that there are advantages to purchasing an NT system over a Linux-based system for real-world file and Web service. Wise system administrators, IS/IT managers, and CIOs should stick with the proven security responsiveness, stability, price/performance, and scalability of Unix-based systems, possibly including Linux-based systems, rather than betting the farm on the Johnny-come-lately Windows NT.