Behind AnandTech - The Search for the Perfect Servers

AnandTech gets the Forums...Servers 3 & 4

That two server setup kept us going for quite a while, until, of course, we decided to start hosting the AnandTech Forums on our own servers. The decision to move the Forums to our own servers was made when we realized that it was time to run the site off of more than just a pair of boxes, so if one box were to go down, the entire site wouldn't go down with it.

These two servers were generally configured like the first pair, except we were using 550MHz Pentium III Xeons, and the new DB server featured 2MB of L2 cache per chip. This time around, Azzo.com was the supplier we used for the parts in the systems, and they all worked flawlessly.

The only other change that we made with these systems is that the second Web Server didn't feature an AMI MegaRAID card, rather an Adaptec AAA-131U2, which is a single channel version of the three channel 133U2 card we used in both DB boxes.

Unfortunately it didn't take long for this four server configuration to become saturated. After examining the cluster's weak points, it was quickly realized that the DB servers weren't having a problem keeping up with the traffic. On the highest traffic days, the DB servers were never showing over 20% usage. It was the Web Servers that were feeling the most pain, they were simply running out of CPU time to process the tasks that remained in their individual queues. Knowing the problem, a solution was pretty easy to find...add more Web Servers, but what kind and how many?

Web Servers are interesting devices, mainly because when you're dealing with a cluster of them, bigger doesn't always mean better. What exactly does that mean? Well, for our situation, we could've added two more dual Xeon boxes and gained a bit of a performance boost, or added 4 or 5 single processor Pentium III boxes and seen an even larger performance increase. The main reason for the latter is two fold; First of all, the Pentium III Xeon is still realistically limited to 550MHz since all of the faster CPUs fail to have any real motherboard support without requiring an RDRAM based platform (i840). If each of these servers were to have over 512MB of memory, obtaining that much RDRAM would be quite costly, and not the best option for us at all. Secondly, when migrating from a single processor to dual processors, you don't see a perfect scale in performance (at least under NT/2K). Instead, adding a second CPU often gives you another 20 - 40% performance increase, definitely not the performance increase you'd see if you were to cluster two single processor systems together.

Now, remember what we mentioned earlier, that for Database Servers you want a decent speed CPU but a large L2 cache but for Web Servers you want something with a fast CPU and a fast L2 cache, but not necessarily a large one. With the Pentium III available in speeds up to 1GHz (realistically speaking we could probably get 933MHz parts at a better value) and featuring a full speed L2 cache, there was quite a good chance that a single 1GHz Pentium III server would be able to outperform a dual 550MHz Pentium III Xeon server in our type of Web Server applications. So we had our problem solved, except for one big part of the equation we hadn't considered, what platform would we run these 133MHz FSB Pentium IIIs on?

The i820E chipset is a definite-no, for the reason we mentioned above: RDRAM is still not a cost effective solution for servers. The BX/GX chipsets were a possibility however there is absolutely no way we were going to overclock our servers to run at the 133MHz FSB, meaning we'd be limited to Pentium III 850s but more importantly we'd be limited to the 100MHz FSB and a 100MHz memory bus which could severely hinder performance.

The i815E chipset was the next consideration that came to mind, the only problem here was that the i815E chipset features a 512MB memory limit which isn't acceptable for servers that constantly have a 300MB memory usage under normal load, a number that can most definitely peak above that.

This left us with the VIA Apollo Pro 133A chipset as the only remaining option, however we were talking about a chipset that hadn't really been intended for a high-end application such as web serving. In bringing up the 133A chipset we were struck by an even better idea, what about VIA's KT133 chipset as a platform for web servers...?