Slashdot Mirror
How to Build a Fad Website: AmIHotOrNot
webword writes: "Web Techniques is running an interesting article written by James Hong, one of the masterminds behind AmIHotOrNot.com (now known as HOT or NOT?). Before you decide to skip over this, consider that Hong and company used Apache, PHP and MySQL to build their site. They found that these open source tools ran much better on a 700-MHz Pentium III than a quad processor Sun E220. Hong also covers their moderation system, advertising arrangements, and how they were able to scale to handle 1.8 million page views per day after being in operation for a mere 8 days."
This gets to the fundamental thing people don't understand about measuring the abilities of computers. There are two ways to measure the capability of a systems: speed, and power. They are related but not closely. See, a 5-year-old SparcCenter 2000E with 20 quite slow CPUs and a few gigs of memory is more powerful than the latest and greatest dual 1.5 GHz Pentium 4 peecee and will almost certainly remain more powerful than even the most expensive and fancy peecee for many years to come.
More powerful. Not necessarily faster. Power measures how much work it is possible for the machine to do in a given amount of time. Power is a general measure that assumes you are running a wide variety of different kinds of jobs, or perhaps a large number of similar ones. When measuring power, you must assume that the load of a system should be at least 0.9 per CPU all the time, and that large amounts of data may need to be moved around. Moving data and sustaining high loads for long periods of time without degradation of responsiveness is an indication of power. That SC2000E can serve many millions of pages a day without breaking a sweat. That's power.
Speed, on the other hand, is a concrete measure of the time it takes to run a certain single linear task, such as a kernel compilation or serving a single page to a single client. This measure rarely places any premium on scalability or the ability to move data. Instead, this measure is typically dependent on the speed of the single fastest CPU in the system. Peecees have great speed - the CPUs are clocked very high and execute very complex instructions. Sun systems do not generally fare very well in this area, especially compared to cost.
The design goals of a Sun are evident from the specifications of the systems - buses are wide, not fast. Latencies can be quite high. CPUs are clocked fairly low and execute simple instructions. The system is designed to allow tremendous throughput - power. Load up on memory, disk controllers and storage, and CPUs, and the huge buses will deliver data well as the load rises. For a web server, for example, serving a single page to a single client is fairly slow compared to a machine with greater speed - remember, a single linear task. But serving one more page beyond the 5000 already being served is where the Sun will really shine. The peecee's memory bus will be saturated quickly; it is clocked fast - much faster than the memory itself in most cases - but is fairly narrow, and the system buses contain bottlenecks. The system was designed for speed, not power. The Sun, meanwhile, can serve the 5001st page nearly as quickly as the 1st. That is the difference between speed and power. The inability or unwillingness to understand the difference between the two is the *only* reason there are peecees in server rooms.
Every system has certain design goals. Peecees and peecee-like systems are designed for single linear tasks - the kind you find desktop computers generally do. Because people - especially traditional nontechnical peecee users - generally do not multitask very well, there is no sense designing their architecture for power. Instead, designing for speed makes more sense. In the server room, however, the exact opposite is true. Unless you're expecting very light use of services, the speed of a system is meaningless; power should be the main design criterion. Using an architecture designed for a completely different problem space just because it happens to look cheap (I won't even start on TCO issues with peecee servers...) is entirely inappropriate. I would fire a sysadmin who recommended peecees for a server environment.
and solaris is a bloated pig.
Yep. For systems with four or fewer CPUs I generally recommend linux. Linux on Suns is as stable as solaris if not more so, and the scalability issues don't generally come into play until 6-8 CPUs or so (with recent kernels). For larger configurations, the very design characteristics that make solaris such a dog on smaller boxes make it outperform linux. It all depends on the goals of the system. In most cases, a larger number of smaller systems is more reliable, less expensive, and more responsive than a small number of large systems. Thus, for example, a web server farm might be better designed as a load balancer and 16 Sun 420s running Linux than as a single E6500 running solaris. But that's another issue altogether...
For the href paranoid: http://www.amiallyourbaseornot.com/?pic=CABG
I almost fell out of my chair when I saw:
Free Online Woodworking Resources Directory
So are the other 90% immortal? =]
No such creature, I'm afraid.
However, HotOrNot is a "context free" metric. You look at a single picture and decide that the person is hot or not. Unfortunately, this isn't all that useful as the answers tend to be very close to either "1" or "10" A much better implementation would be just "thumbs up" or "thumbs down." K5 also suffers from this problem when it asks users to rate comments on a scale of 1 to 5. Keeping it simple would make the ratings much more effective.
Pick the Hottie on the other hand implements a contextual metric. Instead of rating a picture on it's own you look at two pictures and click the one you think is hotter. It's much easier to decide between coffee or tea than it is to rate coffee on a scale of 1 to 10. Effectively the site is sorting pictures using human judgement for the comparison function. This way you get much more useful results. With this system you can get the "Top 10 Hotties." With HotOrNot there are probably thousands of images that are 10.0 or 9.9.
Justin Chapweske of Open Cola gets the credit for pointing out this one.
burris
A great body: +3
Showing some cleavage: +4
Voting your ex-girlfriend a 1: Priceless
There are some things money can't buy...
--
--
We have fought the AC's, and they have won.
One of those is more shallow than the other? They are both one-sided representations of people, and both are poor substitutes for actual human interaction. They are simply aimed at different demographics.
I don't see an enormous difference between posting this on Slashdot and then waiting to see how it gets rated, and posting a picture of myself on RankPeople. You're just getting ratings on different aspects of yourself, that's all.
-Keslin, the naked nerd girl
-Keslin, the naked nerd girl
Simple. None of the images are hosted on their site. They encourage users to sign up for dummy accounts with Geocities and such to host the jpg, and foot the bill for the bandwidth. Each HotorNot pageview only amounts to a 4 or 5k script for formatting and to access the DB. Contrast that with the fact that when slashdot serves a comments page, it's pulling several hundred k from the MySQL database and over the 'net.
When you think about, HotorNot's scheme is quite brilliant. A little shady, but brilliant.
Slashdot: Open Source, Closed Minds.